Intel processors crashing Unreal engine games (and others)

adamge said:

I returned to this thread and saw a lack of activity. "Is this whole thing a nothingburger," I thought?

But the fruit exasperation is real. Something is afoot here. I'll keep an eye on it.

Click to expand...

igor_kavinski said:

But the fruits! The coffee!

Do you know what kind of actions these are?

These are the actions of worthless management refusing to take responsibility for their bad decisions and instead, punishing the worker bees who were only doing what they were told anyway. Now they are making the situation worse by pissing off their workforce who will probably now resort to doing only the bare minimum and not stress it because no one knows who's gonna get fired so why waste time and energy stressing themselves out? Just check in, bide your time, do what's required and get out. No more enthusiasm to show off their talent or good thinking skills.

I'm currently in that mode too coz of what management did to me over the years. I've been picking a lot of fights lately with whoever I think is favored unjustifiably by management, in the hope that they will finally get fed up and fire me so I can move on to doing something better with my life.

Click to expand...

Steltek said:

Yeah, maybe the fruit thing is why that group of senior Intel CPU architects left a week or two ago to start their own company.....

Click to expand...

Steltek said:

Yeah, maybe the fruit thing is why that bunch of senior Intel CPU architects left a week or two ago to start their own company.....

Click to expand...

DAPUNISHER said:

While AMD is fighting them in data center, the fight with ARM in mobile may be the harder one. It's most of the PC market, and ARM may end up doing what AMD never could.

Click to expand...

DAPUNISHER said:

Nothingburger? If you plan to astroturf and gaslight it won't end well.

Click to expand...

adamge said:

Fair.
I shouldn't imply the problem is not a problem.
But its impact appears to be well contained from a business and greater market perspective. This won't bring down Intel.

Click to expand...

gdansk said:

AMD survived spinning off its fabs. AMD survived Bulldozer. AMD survived selling and leasing back their own headquarters. AMD survived having a market cap of $2B.

Intel isn't at that level. They'll be fine even if they are struggling.

Click to expand...

igor_kavinski said:

I got super close to getting the ASUS SD X 78 OLED 32GB/1TB laptop for $1200 on a $100 monthly installment today, with my cc details entered and installment plan selected. Only thing left to do was clicking the order button. Stopped at the last minute coz I didn't want to receive the laptop at work and it didn't seem like I would be able to get a day off to receive the laptop at home. Later on, I felt better that I didn't order it. It's gonna get discounted even more and who knows if Lunar Lake or Strix Halo turn out to be really ground breaking. But I was so tempted. Oh so tempted!

Click to expand...

Steltek said:

I was reading something somewhere today that Intel has a "Arrow Lake Halo" enthusiast mobile workstation CPU sku coming out with a beefed up GPU tile to compete with Strix Halo. Don't remember where I saw it, though. Someone in China reports it being listed on a shipping manifest there.

Click to expand...

maddie said:

You can't be serious? IBM, Xerox, sound familiar? Do you outright reject such a possibility?

They will be Intel in name only. The present company has to die and be reborn, same as happened to AMD, in order to survive. If they don't, they slide into irrelevancy, as business as usual can't work.

Click to expand...

gorobei said:

amd q2 earnings doubling (y/y) in server was a major signal

Click to expand...

gdansk said:

Hmm, I thought nearly a billion of AMD's DC revenue was MI300 not Epyc. That would be perhaps 2/3rd of their DC revenue increase.

Click to expand...

gdansk said:

AMD survived spinning off its fabs. AMD survived Bulldozer. AMD survived selling and leasing back their own headquarters. AMD survived having a market cap of $2B.

Intel isn't at that level. They'll be fine even if they are struggling.

Click to expand...

igor_kavinski said:

But the fruits! The coffee!

Do you know what kind of actions these are?

These are the actions of worthless management refusing to take responsibility for their bad decisions and instead, punishing the worker bees who were only doing what they were told anyway. Now they are making the situation worse by pissing off their workforce who will probably now resort to doing only the bare minimum and not stress it because no one knows who's gonna get fired so why waste time and energy stressing themselves out? Just check in, bide your time, do what's required and get out. No more enthusiasm to show off their talent or good thinking skills.

I'm currently in that mode too coz of what management did to me over the years. I've been picking a lot of fights lately with whoever I think is favored unjustifiably by management, in the hope that they will finally get fed up and fire me so I can move on to doing something better with my life.

Click to expand...

moinmoin said:

Yeah, the big difference between AMD and Intel is going to be working atmosphere. At its lowest point AMD still reportedly had a great working atmosphere despite low pay and unsure company future. Intel is well in progress of destroying said working atmosphere even before it reaches its low point. That's why fruits matters, actions like these set the tone.

Click to expand...

Nothingness said:

Is there some "trustable information" about how much engineers get paid in these companies? No, I don't trust Glassdoor.

I know Intel has the reputation of being on the lower end. And in the recent years, they've become an investor company (they're hopefully done with that); I'm glad I didn't pursue with a job offer there 15 years ago. And that fruit story is so ridiculous; they'd better fire their marketing droids and get back to engineering, that'd save much more money. (I have much more to say about Intel as a company, but I already ranted before, no need to repeat again and again.)

Click to expand...

coercitiv said:

Oh Buildzoid, why do you have to do this? Another 1h40m video I HAVE to watch because I want to know about this.

I guess no movie tonight for me

Click to expand...

igor_kavinski said:

Spoiler: TLDW

Hey everyone, Buildzoid here. Today, we're diving into how load lines operate on the Intel LGA 1700 socket, specifically addressing why Intel introduced a voltage limiter in the 129 microcode that prevents CPUs from requesting more than 1.55 volts from the motherboard.

This video is quite detailed as we’ll discuss the fundamentals behind voltage regulators and their functionality. We’re using a Z790 AORUS Pro motherboard for this demonstration. The Vcore VRM, or Voltage Regulation Module, comprises various phases and is controlled by a specific chip that manages the output voltage by sending a PWM signal to the power stage. This signal turns the power stage on and off; the width of the signal determines the output voltage.

Importantly, the controller must accurately measure the output voltage and relies on Intel-provided pins directly on the CPU for this purpose. The voltage regulator measures voltage at the CPU, not on the motherboard. This distinction is crucial because if it only measured on the motherboard, it wouldn't provide reliable voltage readings due to drop-offs occurring through various connections leading to the CPU.

Intel includes both positive and negative pins for voltage measurement. This dual pin setup allows the regulator to measure voltage as a difference between the two points—essential for measuring accurately because voltage is always relative. If only one pin were provided, fluctuations due to high current flow could lead to inaccurate readings that could cause the voltage controller to miscalculate the voltage supplied to the CPU, potentially resulting in damage.

Additionally, the controller uses a feedback mechanism to monitor the CPU voltage, which does not often provide accurate readings via monitoring software due to how sensitive these connections are and their placement on the motherboard.

If the feedback connection were to fail, there are backup circuits, usually involving resistors, that can still provide an approximate voltage reading, although not accurately. The voltage regulator can still sense voltage presence though it would be inaccurate, mitigating potential issues of the controller misreading zero volts and causing excessive voltage output.

The controller gets voltage requests from the CPU via the SVID interface, which directly specifies the voltage required. Without load, this is a straightforward translation between PWM signals and output voltage. However, imparting load introduces resistance—every connector, socket, and PCB trace adds resistance, leading to voltage drop-offs. For instance, higher current could create significant under-voting at the CPU.

In a steady state with no current, the voltage reading would match the request. But with practical loads, fluctuations occur due to the resistance encountered in the circuit. For example, at 10 amps, one milliohm of resistance might drop the voltage significantly, explaining why multiple phases in the VRM might show slightly different output voltages under load.

This effect becomes critical when a CPU suddenly swings from low to high current demand (like from idle to load), which could let the CPU dip below operational voltage thresholds, leading to instability or crashes.

To address these issues more efficiently, load lines are implemented to prevent the CPU from exceeding voltage specifications under heavy load. A load line allows for voltage sag during heavy CPU demands, effectively tuning the voltage based on system load. For older CPU architectures, this was more manageable; however, with modern CPUs like the i9-14900K, the voltage droop becomes a critical concern.

With the load line constraints set to 1.1 milliohm, the CPU requests voltage—let's say for maximum performance—often result in an output that’s too high, potentially hitting 1.6 volts under certain conditions, which poses risks of degradation and damage.

The aftermath of this leads to discussions around system-level efficiency, optimization of voltage requirements, and ultimately reactions to power management issues on new and existing motherboards. If manufacturers don’t consider these parameters, they could be setting themselves up for failures.

Intel's limitations on voltage requests and tweaks with microcode represent attempts to safeguard their CPUs from the very situation we discussed. The design of the load line slope system adjusting dynamically based on CPU loading is critical to preventing sustained operation in unsafe voltage conditions, while also exerting cooling and energy efficiency.

Overall, understanding these mechanisms will help inform how processors respond to loads, why they operate the way they do under stress, and highlight Intel's efforts to try and reign in erratic performance under potential high-speed conditions.

Thanks for tuning in, and I hope this breakdown clarifies how load lines function within the context of the Intel LGA 1700 socket. If there are any lingering questions or overlooked details, feel free to ask!

Spoiler: FULL TRANSCRIPT

hey guys buildzoid here and today we're going to be taking a look at how load lines work on the Intel LGA 1700 socket
and basically why Intel had to add a voltage limiter with the 129 micro code
that stops the CPU from requesting more than 1.55 volts from the motherboard so
uh let's get into it um now this is going to be really long because
uh we need to basically go through how voltage Regulators work um in terms of
the how like well some of the functionality of the voltage regulator so anyway this is a z790 orus prox
motherboard um the Vore vrm on this board well technically it should be just V cor VR um instead of the m stands for
module so and and it's not you know it's not a module that like slots into the board but anyway so the vcore regulator
so VR um is this set of phases over here
uh and the controller for those um is this chip over here right and the way
the controller uh controls the output voltage of all of the phases um is that
it sends a pwm signal to the power stage and that just switches the power stage
on and off and depending on the pulse width of the pwm signal you get more or less voltage right um so that's that's
like the basics of of how our controller works now obviously uh the controller
needs to measure the voltage that is coming out of the phases uh and it needs
to measure it somewhere um right and depending on where it measures it things
can get like you're you're going to like it's either going to work or it's really not going to work um so because of that
uh Intel actually provides a dedicated uh measurement connection for the CPU
so this is the first thing I want to make absolutely crystal clear because I
see a lot of people just not understand this um the voltage regulator measures
the voltage at the CPU it doesn't measure it on the motherboard it
measures it at the CPU through pins provided by Intel to get a accurate
voltage measurement of the voltage at the CPU okay
it measures at the CPU it regulates the voltage uh at the CPU it does not care
about the voltage on the motherboard cuz it Doesn't measure that it has literally no idea what the voltage at each
individual phase is because that's pointless okay it doesn't need to know that it just needs to know what the
voltage at the CPU is so Intel provides these two pins right here to do that and
it provides two and they provide two pins because when when you have lots of current flowing uh you need to measure
both the like you you you um ah hell okay let's start over they provide two
pins because voltage measurements are always relative between two points right
so a single point in space does not have a voltage there is all like if you have two points then you can measure the
voltage between the two two points so Intel has to provide both a negative pin
and a positive pin so that you can measure between them um and you can basically think of these as one end is
VC cor and the other one is ground and this is really important that they provide both a negative and a positive
because uh if they only provided one pin um the amount of current that is Flowing
right cuz current always Flows In A Circle right like it always has to
return so if you have 200 amps going into the CPU you also have 200 amps
going back out of the CPU and into the voltage regulator right like the so you
have this like 200 amps of current flowing which means that there is a voltage drop between the lowside mosfet
of the which is the mosfet that like handles the return current um and uh the
CPU silicon like there's a like if you measured um now unfortunate oh I guess
we could just well so if you found found out which side of one of these capacitors is negative right um so let's
just say it's this side I don't know I didn't bother to check uh it doesn't really matter but like if you measured
from the negative over here versus say the uh actually this is probably the
ground pad or yeah this is probably ground um so if
you measured from like here uh to the ground of the CPU you would actually see
a pretty substantial voltage drop because of all of that current that is flowing back from the CPU and towards
the voltage regulator which means if intel only provided one pin uh for the voltage measurements the like the
voltage controller if it was refer like measuring the voltage at the CPU relative to its own ground uh it would
measure a higher voltage than what the CPU is actually getting um so that's why
Intel provides two pins not one because if they provide or actually why they provide any pins at all because if you
didn't have any pins then uh you would actually just have no idea what voltage the CPU is getting whatsoever um it
would be just completely all over the place um so um yeah so the voltage
regulator has a uh measurement connection
directly to the CPU and we'll call this the feedback um slash uh
s like you'll see it's sometimes called vcore sense um annoyingly enough some motherboards get really weird about the
like labeling things Vore sense especially MSI has like a VCC sense thing that doesn't measure correctly
but if motherboards were like you know like I I just it motherboards man
like like this in this case like the the issues that is dealing with with the
excessive voltages that the CPUs are getting that's not the motherboard's fault but like the fact that if you fire
up a like Hardware info on a bunch like on a bunch of LGA 1700 motherboards
getting accurate voltage measurements is really hard because the motherboards don't give you access to the
measurements from the voltage controller they just don't and this is for a good reason cuz this uh feedback circuit is
very sensitive so you don't really like the chip that most of your monitoring
software actually pulls up is the super IO chip and the super iio chip lives like somewhere down there on the
motherboard so you don't really want to go grab your very very sensitive feedback circuitry you know sense wires
and run them over to the super iio that's on like the bottom bottom edge of the board somewhere uh that would be
pretty bad because if you run that wire past like um I don't know some like if
you run it past like a voltage regulator or something uh you're going to have a bunch of junk in your in your you know
feedback circuitry and then then your vrm isn't going to work correctly so uh it's for a good reason that this circuit
like this connection is like exclusively used for the controller um and there are
some high-end motherboards like this one that actually do provide like access to the the feedback like the voree sense
measurements um but they require extra circuitry cuz the there's no like
basically uh there's no requirement for the motherboard to provide you with accurate software voltage readings like
that's just not a thing like nobody cares so here we are but yeah High some
high-end motherboards will basically give you a op amp so that's what this chip is and all this chip does is it
copies the uh voltage from the feedback circuitry that's going to the uh vrm
controller and then it outputs that over to the super iio and this is fine because even if there's a bunch of
garbage on the like super iio side um of the connection it doesn't get fed back
into the feedback circuitry right so the feedback circuitry is not uh like is
still like safe from whatever is going on with the wires connecting to the the super IO um and gigabyte does this on
some of their boards Asus does this on a bunch of their boards MSI doesn't do this ever they have like
a uh you know what not going to bother being nice they have like a half-ass implementation where I think they only
take like the positive end of the feedback circuit and send that to the super iio so it reads like 30 molts high
at 100% load but um oh well you know and ack uh I don't think does this
at all either um so yeah um though if you're lucky on some other boards you
can just directly pull the voltage controller but that doesn't always work either um so
so yeah um cuz like some some motherboards just like straight up block access to the controller itself uh and
then a lot of the time the controller just isn't supported by the monitoring software so you can't you know get a reading from the controller
anyway but uh where was I right so we have this feedback circuit and that's what the controller uses for monitoring
the voltage at the CPU and unfortunately you probably don't have access to the uh very accurate High Precision voltage
measurements that the controller has because that's just not a part of the design
concerns of the LGA 1700 platform um now you might also be wondering well what if
these two pins break wouldn't that be really bad yes yes it would be really bad and so there is a fallback
connection uh typically that's implemented in the form of a like pair of resistors um that'll just connect
somewhere on the motherboard uh usually inside like the CPU socket so if you uh
look over here so like that that connection will be made into somewhere in this area um right um and the point
of this fallback connection is like the way this is set up is usually you'll have like 100 ohm
resistors um and the whole idea is that if these two pins get disconnected then
you still get a voltage measurement um from the mother board right but it's not
accurate but at least you don't end up in a situation like cuz if this didn't exist and your pins broke uh what would
happen is the vrm would fire up and the voltage at the pins would like they'd just be floating right so they might be
at Zer volts they might be at something really high um and if they're at Zer volts then the voltage controller is
like well the voltage is zero so obviously I need to increase the pwm signal so the voltage on the output
would progress like just shoot upwards right cuz the controller is like I'm raising the pwm signal and the output
voltage is still zero and it's still Z and it's still zero and luckily uh most
modern high-end controllers do have a Max pwm width so they'll just like if
they hit like 30% or something they'll just cut out because something's evidently wrong um actually I don't
think it's 30% but there is like a max pwm pulse width at which they will just
shut down um so even if the the connection is gone your output voltage won't just like shoot past like 3 volts
or something but two volts is already plenty to like more than enough to destroy a CPU um so you do need this
like backup circuitry of like these 100 ohm resistors so that if the pins to the CPU break uh the voltage controller at
at least can still see what the output voltage of the regulator is right um and the reason these are like 100 ohm
resistors is uh the connection to the CPU is lower resistance than these resistors so if the CPU like if there's
nothing wrong with the the socket then the voltage uh from the CPU overrides um
the voltage that's sort of coming into the circuit from the uh through the resistors right so that's what's going
on with that and for the purposes of this video you can basically forget that that like backup circuitry like backup
feedback circuitry exists um but I've felt like it's necessary to point that
out because yeah it it does exist um your motherboard like your your voltage controller if something goes wrong with
the connection to the CPU doesn't just blow up your CPU
okay um at least if the motherboard is designed properly I've never seen a motherboard that doesn't have the
fallback connection but like I don't know technically it's not like it's
probably required but yeah like you know I like I guess if you like shaved the
like if you had a banged up secondhand motherboard right maybe they sh like somebody shaved off the resistors and
then they bent the pins and then yeah then you'd be screwed but generally speaking uh not a concern so anyway the
controller monitors the voltage at the CPU like at the Silicon
okay um so um the CPU uh tells the controller what voltage
it wants over the svid interface so we're just going to draw that out over
here all right so this is just a digital output um the s in SVD stands for serial
um so it basically just sends a string of ones and zeros uh other vid standards
are things like you have PID which is parallel vid where you'd basically instead of having like one well svid has
like one data line and then there's a clock line right um but PID you'd
basically had like seven or eight pins on the controller and you would encode the voltage by pulling pins high or low
uh then there's pwm vid which is what Nvidia uses where they just use a pwm signal to the controller and then the
controller converts that pwm signal into a voltage level um so you know there's a
bunch of different ways but Intel uses svid where they just send a string of ones and zeros and a clock signal uh on
a second wire so that the controller can decode that into voltage requests so if
our CPU sends a svid request for 1.2 volts and we have uh like and we don't
have a load line right um then the controller is just going to look at the
voltage at the end of the sense circuitry right it's going to look at the voltage like I don't like I don't
actually know where those pins connect on the CPU but probably pretty close to the Silicon cuz it's worth pointing out
uh just uh like if you look at the pin out here like these red pins these are all VOR and now I haven't checked this
for LGA 1700 but my past experience is with other Intel CPUs is that just from
like here to here you already have a significant voltage drop on the CPU
substrate itself right so if you like so that connection has to go to the Silicon
like it can't just connect to some random like it needs to go pretty close to the CPU it can't just go to some
random place on the substrate because uh you'll notice also this one's upside
down right corner are our corner so
um wait if that's now I'm confused and we'll just
pretend that VOR comes in over here so like you know if you measured the voltage over here that's going to be different than the voltage over here
that's going to be different than the voltage over here that's going to be different than the voltage on these capacitors so uh yeah uh I don't know
specifically where exactly the voltage regulator sees the voltage on the CPU but it is wherever Intel intends for it
to be measured so uh we'll say it's somewhere near the center um anyway so
the controller goes like okay I need to get 1.2 volts over here and it just sends a pwm signal to the power stages
that leads to the voltage over here being 1.2 volts now why do I say leads
to the voltage at the CPU being 1.2 volts well um if we don't have any
current flowing right if we have zero amps then this is pretty simple you just 1.2 vol at the inductors means 1.2 volts
at the CPU so in that case we can just send a pwm signal of 10% and all is well
right you because we have 12 volts coming into the power stages and we're just going to pretend that it's 12 just
straight 12 volts right uh um so we have 12 volts coming in 10%
pwm signal 1.2 volts comes out we get 1.2 volts at the CPU all is well but
this is with no current flowing so this isn't realistic cuz CPUs you know like current flows through CPUs at least if
they're actually doing something um so let's say we're outputting um like uh 10
amps well if we're outputting like 10 amps and the uh motherboard um then like
obviously the issue we run into is that the motherboard and the CPU socket and the CPU substrate they all have
resistance right um and from my personal experience the
socket itself seems to have about like based on like measurements that I've taken the socket for LGA 1700 seems to
be about 0.2 milliohms all on its own so that's before you even account for the
motherboard um and then for the power plane of the motherboard itself right
cuz you basically need to somehow like this is an eight layer board so technically there could be a connection
on an internal layer that goes over here right this is a bunch of eor pins over here um so like technically there could
be a connection there but if you have a cheaper board that's say a six lay PCB
uh yeah the power from the phases down here probably just has to do this um and same for like you know so
all of the current ends up having to squeeze through like the the top corner of the corner of the CPU socket which
leads to some funny funny uh thermal situations if you have a really crappy
board where like this area of the board gets really really hot um cuz all of the current has to squeeze through there to
get into the CPU socket which is like great um and anyway so yeah the motherboard has some resistance the
socket has some resistance the CPU substrate has resistance which is why we have that connection all the way to the
CPU um so you know let's just say all of that adds up to a resistance of 1
milliohm um well if we're putting 10 amp s through a resistance of 1 mohm the voltage over on these
inductors uh like we're going to have like 10 MTS of voltage drop right so the voltage on our inductors has to be like
1.21 volts but like I drew earlier this inductor
down here right this phase down here is significantly further away from this corner of the socket than say this phase
so in reality um the voltage over here might be like 1.2
uh one one here it might be like 1. 1209 and you get the idea so every phase like
on a running motherboard every single phase will actually have a slightly different output
voltage which is why the controller doesn't measure the voltage
on the motherboard because that would just make like it would be a horrible mess every motherboard would like do
different things um and then depending on how good your socket contact right
like imagine if the if like you have less or more mounting pressure and that affects the the contact with the pins uh
for for the CPU it's like yeah no this this isn't going to work so that's why
the controller measures the voltage at the CPU um and then it just and the
funny thing is like it's not even the controller technically it can uh try to balance the current through each of the
phases CU it does measure the current flowing through each phase but it doesn't know the output voltage for each phase so the controller can do things
like send a slightly shorter pwm uh pulse width to the phases that are
outputting more current uh in order to push some of that current over to the other phases that are working less hard
and this is like a efficiency optimization primarily um is making sure that the current really is evenly spread
across all of the phases um though some controllers will also do some like I
think Asus does this on some of their motherboards where they'll actually uh spread the current based on operating tempat temperature not balancing um
which could actually lead to better efficiency than the like e like the even current spread uh setup but anyway so
the point is like the voltage at each of the inductors is going to be slightly different and the controller neither knows nor cares cuz it measures the
voltage at the CPU and as long as like CPU asks for 1.2 volts CPU gets 1.2
volts what the output voltage from the inductors is doesn't matter right
literally controller doesn't care um now if we're outputting 100 amps uh things
get uh well more complicated because at 100 amps 1 milliohm of resistance means
that uh we're just going to pretend that our vrm has now turned into a point source of power instead of what it
actually is which is a bunch of different phases uh and so on average the vrm will basically end up at like
1.3 volts now right because we have that 1 milliohm of
resistance now again in reality you would end up with like the really far away phases sitting significantly higher
voltage and the nearby phases sitting at lower voltage but point is um yeah so
you know 1.3 volts over here but the CPU is still about 1.2 volts right cuz the
controller measures the voltage of the CPU it doesn't care that there's 1 milliohm of resistance it has no idea
there could be 2 milliohms and then it would just set this to 1.4 and it still wouldn't care right cuz it measures the
voltage at the CPU um so um and obviously if we keep going
up in output current like if we hit 300 amps right then the voltage at the the vrm is going to be like 1.5 volts cuz
300 amps through 1 milliohm uh the pwm signal width is going to be about
12.5% um and the vol at the CPU is still 1.2 Vols it's great right like the
controller just deals with this so why do we need load lines well the thing is
all of this that I've described here is for steady state so that means the current isn't changing right if the CPU
is continuously pulling 300 amps then yeah the voltage at the inductors is
going to be something around 1.5 volts and the voltage at the CPU is 1.2 volts
but what happens if the CPU suddenly goes from 300 amps to
zero all of that voltage drop that we have through the power plane because of
our high current draw just disappears right and to make things
worse uh we have a bunch of excess energy stored in the inductors um because the current flowing
like inductors store energy um based on how much current is flowing through them so if you want to quickly go from having
a lot of current flowing through an inductor to having less current flowing through the inductor that energy in the
inductor needs to go somewhere um so on A3 like so when the CPU transitions from
like 300 amps of you know like steady state to zero um yeah the the voltage suddenly
like it goes from 1.2 volts to basically like 1.5 or higher
it just shoots through the roof right because you have like first of all the board is like charged to like 1.5 volts
anyway right these capacitors are literally right on top of the inductors um and then of course you have all the
extra energy stored in the inductors themselves and you basically end up with a nice big fat voltage Spike that goes over 1.5 volts um not ideal
so um but ultimately like this isn't really like you know as long as the
spike isn't too too big it's not actually going to damage the the CPU so this isn't too concerning but the bigger
problem is the opposite transition where you go from zero amps to
300 right cuz if the if the output voltage or let's say from 10 amps to 300
right um because in that scenario
the um volt like at zero amps right the vrm is like at .2 Vols and here is at
1.2 Vols but this resistance exists so if we suddenly ramp up to 300 amps the
voltage at the CPU is just going to fall off a cliff right and the controller also
doesn't react instantly and the power plane does have some inductance to it and the inductors have a lot of
inductance to them so you end up in a situation where you know if you're at zero amps and then suddenly the CPU
jumps to 300 amps uh the volt temporarily falls off a
cliff right and at some point the controller is like wait a minute this isn't uh 1.2 volts this is 1.18 I need
to do something about it so it starts turning on the phases um so most modern
controllers can actually turn on multiple F phases at the same time if it's a smaller load step it might just
increase the pwm signal width right or if the if the transition is slower like if you go from Z amps to 300 amps very
slowly uh the controller might just slow like progressively increase increase the pwm signal width and no big deal right
but uh if you have a really fast transient um where you know the current drw jumps in the span of like a fraction
of a actually like not even a fraction of a millisecond we're talking in like the range of micros seconds right that
the the current ramps up from like 0 to 300 amps uh yeah the voltage is kind of nose Dives um and then the controller
eventually like catches it and manages to pull the voltage back up to where it's supposed to be um but if you're a
CPU manufacturer and you know you have a Target operating voltage of 1.2 volts
right so it's like oh I want 1.2 volts in order to run 5 GHz and your CPU is happily running you
know like here it can run 5 gahz um and here you can run 5
GHz uh that's not a g um so here you can run 5 GHz but the bottom of this might
be like I don't know 1 uh one volts well this will not run 5 GHz this
will run maybe 4.8 or 4.9 or something right obviously
like depends on the Silicon um so if you have a chip running at 5 GHz and your
voltage suddenly does this uh your chip crashes is what happens um unless you have some kind of
fancy clock stretching mechanism like what AMD has or what Nvidia has or even Intel at this point like as far as I can
tell they do have clock stretching now um but anyway um if you don't have clock
stretching like say when load lines were first implemented yeah this will actually just like crash your CPU so
this is not good um and so CPU manufacturers and like voltage regulator
uh manufacturers uh came up with the idea of load lines uh which don't really solve the problem of like oh the voltage
jumps off a cliff but what does solve is like the the thing is okay well let's
bring this all back um like it doesn't really solve the fact that like the voltage just inherently
wants to fall like fall down to 1.1 volts um but pulling the voltage back up
to 1.2 volts if you're limited to 4.8 GHz due to the voltage
regulation right like you can't you like if this happens and you want your chip
to be stable you have to be like okay it's actually 4.8 gigahertz at uh 1.2 volts right then this is like a
massive waste of energy because you have all of this voltage over here that you can't actually use cuz if you clock the
chip to 5 GHz and there's a sudden transient it crashes but um if you clock
the chip to 4.8 GHz well you don't need 1.2 volts but you also can't just drop
the like you can't you know lower the chip to 1.1 volts 4.8 um and then 1.1 Vols 4.8 because if
you start at 1.1 volts right your voltage Dives again and ends up at 1 volt now so load lines were initially
implemented as a uh efficiency optimization basically um cuz yeah if you're like
your voltage regulator just it can't maintain perfectly flat voltage it's physically impossible even like for you
have the delay of the like first of all the controller like the controll needs to notice that the voltage is dropping
right and that doesn't happen instantly the controller isn't instant it needs like it takes time to measure it it
takes time to adjust the pwm signal widths uh it takes time for the mosfets
in the uh Power stages to turn on it takes time for the current through the inductor to ramp up uh and then your
power plane itself has some amount of inductance as well as the pins and the substrate so like even like even if you
removed the like measurement delay the voltage would still fall off a cliff maybe not as much but it would still
fall off a cliff um so um at some point
CPU designers and uh voltage regulator designers decided Well okay um if if we
can't you know like if we have these massive just undershoot spikes that cause
instability there's no point pulling the voltage back up cuz we can't clock the chip any higher even if the voltage
recovers eventually um so instead they added load
lines which basically just let the vrm SAG the voltage based on output current
so in this scenario of like uh you know we have this setup here where it's like
okay 1.2 volts at um and the target fre
well like CPU sends an svid request for 1.2 volts um if we tell the controller right that
load line equals
uh this is not going to fit um so we say the load line equals 1
milliohm right now when we go from 1.2 volts the voltage Dives right so now
this is like 1.2 volts at 4.8 GHz so like yeah we are clocking the chip lower
but like you had to do that anyway previously right cuz your other option was like yeah the voltage still hits
this minimum point and then you pull the voltage back up and you have like higher power consumption by 10% actually
actually it's not even 10% if you if you have undershoot like yeah if you have 1
point if you have an average operating voltage of 1.2 volts with an undershoot of
1.1 um you basically have like 20% extra power consumption for no extra stability
in terms of operating frequency um because your a like your average voltage dictates your power consumption
and your unders shoot dictates stability right so if you have 100 molts of undershoot you're literally just like
throwing efficiency out the window um so anyway unless you have a clock
clock stretching me uh mechanism so anyway so here you'd have like 1.2 volts at zero amps and then here you'd have
1.1 volts at 100 amps right because we set our load line to 1 milliohm so the
voltage controller gets a request for 1.2 volt uh it gets a request for 1.2 volts and basically what it does is uh
1.2 Vols minus uh okay 1.2 Vols minus 1
milliohm time uh 0 amps right and some
like that just is zero so 1.2 Volus 0 =
1.2 Vol um over here it does 1.2
Volus 1 * 100 again it's like uh 1
milliohm time 100 amps right and that spits out uh 1.1
volts right and then on the load release the voltage just cleanly recovers back
up to 1.2 volts um so this is basically how load lines work online like Sandy
Bridge CPUs uh LGA 1366 uh bulldozzer
CPUs uh Ivy Bridge Skylake um yeah most of the early
Skylake stuff KB Lake um and um yeah so
this like you know like if you look at this it kind of looks weird cuz it's like but but like your voltage is lower
like why would you want this and it's like well because again like yeah you could pull the voltage
back up but like you still can't run more
than 4.8 gahz so why even bother right like why bother pulling the voltage back
up um also with this load line setup you
can actually get away with less capacitors which is cheaper and we all know that manufacturers love things that
are cheaper and the reason you can get away with less capacitors is because um
especially like the big bulk capacitors these are entirely too far away and too slow to really do anything about this
right here um like through hole aluminum polymers
great for absorbing like big voltage spikes coming from the inductors kind of useless for preventing a brown out on
the CPU right um cuz they're not on the CPU
they're like a mile away behind all of the resistance of the power plane all of the resistance of the pins of the socket
all of the substrate resistance right so like these capacitors over here are basically the only thing that really
affects how quickly This falls off a cliff um so yeah
um light life is very hard for for CPU power delivery um but if you have a load
line right like you don't need as many like you don't need so many bulk capacitors to absorb the like excess
energy from the inductors on the load release because you're now starting at a
lower voltage right so you can like throw a bunch of your bulk capacitors out and it
just makes your boards cheaper so that's cool um so basically load lines in like
the whole point of load lines is like you get better efficiency uh at 100% load not at idle right like obviously
your idle volt to just still 1.2 volts so your idle power consumption hasn't changed but at least when your CPU is
pulling 100 amps uh well it's pulling 100 amps instead of cuz here's the thing if this was 1.2
Vols and your frequency didn't change uh I can tell you from just like from
experiments I've run in the past your current draw is going to be basically 110 uh is actually going to be a bit
more than that um right so I guess if we drew it as like a separate
line I'll do that in red um and we had that Spike right like your options are
basically 4.8 GHz um but with 1.2 volts and like
110 10 amps versus uh 4.8
GHz with 1.1 volts and 100 amps right and notice how like power
consumption is voltage times current here both your voltage and your
current is higher so this is significantly less power efficient than
this right um so that's why load lines were implemented was because like well
the voltage regulator like it is physically impossible to design a power delivery system that isn't going to have
undershoot okay that just it cannot be done there like laws of physics prevent
it from happening um so since you're never going to get
the voltage you want anyway uh you may as well just kind of go with the flow if you will right oh
the voltage wants to jump off a cliff that's fine we're not going to fight it the voltage wants to jump up cool right
and the end result is you get a cheap like you don't need to have so many bulk capacitors your power efficiency is
better uh what's not to like other than the fact that you can't run 5 gahz anymore right so but it's like well
previously you couldn't run 5 GHz anywhere like previously the 5 GHz didn't really work either so you know
not not actually a loss um so that's what load lines do um right is that they
just droop the voltage based on output current and for low current CPUs this is
this is fine this doesn't really cause any major issues because yeah if you have a CPU that you know has a peak
current draw of like 100 amps you just idle it at like 1.3 volts and then you
get your 1.2 volts underload and everything is cool or you idle it at 1 .2 and you get 1.1 underload and
everything's cool but when you design CPUs that have uh you know and this is
like Sandy Bridge like Sandy Bridge LG 1366 like those CPUs had a like
operating voltage of around a vault um maybe 1.2 they pulled like a maximum of
100 amps you know so yeah it was it was cool you had your 1 milliohm you know 1
mohm load line slope the voltage droops by 100 molts
everything's fine no big deal um but then you get to something like say a
14900 K right um and a 14900 K brings some very
real problems to this uh you know I wonder if I could just grab the Eraser
for this what is the Eraser shortcut oh it won't tell me oh yeah it will not tell me what the
Eraser shortcut is technically the back of my P has a eraser but I don't think I have it like configured to work oh why
are you like this
um I wish the P like the Eraser and the paintbrush didn't share brush sizes
that's like so annoying I bet there's some way I can change that anyway so if you have a 14 so that was like an
example of like you know 1.2 volts 100 amps load line just kind of like you get
extra efficiency and everything everything's cool no big deal um but then you get to something like a 14900 K
uh this is a 13900 K but you know 14900 K and the all core
clock right
turbo is like 5.6 GHz and now my CPU is an absolute potato
so it's like .6 GHz at 1.38 but let's say you don't have a potato and it's 1.3 volts so that's not
that high right but then you run into the problem
that a 14900 K has an ICC Max of
307
amps so you know how with 1 mohm of load line slope at 100 amps the voltage
droops by 100 molts yeah with this it droops by like
307 M volts so if you want to get 1.3 volts at
the CPU you need to ask for 1.6 to the voltage
regulator right um and that's bad that's just just
that's just like yep that's not really a good idea is it
um and that's unfortunately exactly what Intel does um and actually it gets worse cuz Intel's like Max right the
max ACL for LGA 1700 man my L's are looking
like C's um is 1.1 mli ohms
right um but now Intel isn't completely stupid so they don't actually idle your
14900 k at uh you know 1.3 volts plus
37 time 1.1 milliohms because yeah that
would be really really dumb that would probably destroy your CPU in a few hours
okay that really wouldn't take very long to take your 14900 K back to being sand
um but um so how does Intel like avoid this issue because at the end of the day
right if if the voltage regulator is set to 1.1 mil
ohms um then in order to get 1.3 volts at the
CPU you need to send a vid request for 1.6 like there's there's nothing else
you can do so how how does Intel avoid that or at least try to avoid that well they look
at how many cores are active and this is why you might notice that in the uh
statements that Intel has released about preventing degradation on their 13th gen and 14th gen CPUs they keep like saying
that you shouldn't use the high performance power plan for Windows because the high performance power plan
uh I guess is like less fast to put the cores to sleep or it might keep them
awake long and basically the the CPU like if like
so if you have one core boosting it doesn't do 300 amps right if you have one core I think it's like like one p
core equals I think about 50 amps um cuz I
don't I don't know what it is for sure two P cores you know it might be
like 100 amps right and so this is like an extra 100 molts uh four P cores and
now now things start get getting you know awkward cuz 4p cores can pull an
incredible amount of current um so I actually don't know what this would be for the like like expected Peak current
but uh let's just say it might be 200 amps right and so on and so on and
eventually you get to the 24 core configuration and the expected current equals 307 amps
actually it's more than that as far as I know but we'll just go like it seems to be pretty close to 307 amps and the
reason I say it's more than that is I actually tested for this with the I5 I didn't test for this with the I I9 cuz
it's uncool um but with the I5 I tested for this and the max vid request sort of
stops changing above 200 like somewhere above 200 amps and the ICC Max for the
I5 is 200 so like going from 200 amp ICC Max uh to 225 there's still a increase
in Max voltage actually you know what I'll just pull that up
um where is
it oh yeah here we go um
so this is from a 13600 k um vrm set to 1.1 milliohms AC load line set to 1.1
Milli OHS and you can see how Peak voltage with 500 amp ICC Max is the same
as it is as 250 amp ICC Max but 4 point like 1.468 for 225 1.43 6 for 200 1.41
for 175 and you'll notice that like this step right here is about 25
molts right 1.4 36 from
1.4116 and we see about another like what is it um 22 M Wait no that's like
32 M volts more Peak voltage so there might be some like extra overshoot getting mixed in from the vrm um but as
you lower the ICC Max the max voltage gets progressively lower um and then you
see basically the same thing for the different load lines I tested also 0.55
and you'll notice that like yeah if you set the lower AC load line you get lower Peak voltages um cuz
uh like Intel like and here's the thing I just I I'm you know like I don't see
another way of dealing with this either cuz if you don't have clock stretching
um then yeah like if you want like if the board has one milliohm of like
impedance and the voltage regulator is set and even if the voltage regulator isn't like it doesn't actually matter if
the board has 1 M ohm or doesn't right cuz let's say um let's say I have a um
let's say I have a 0. 5 mohm motherboard right and the voltage regulator set to
1.1 mohms well the voltage regulator measures the voltage over here right so the target voltage for the
voltage regulator if I say you know I send a vid request for 1 Point um 1.6
volts right and I send that down my svid interface um and there's 300 amps
flowing uh from the mother board then you know our 1.6 volts gets to the
voltage regulator and the voltage regulator does 1.6 volts minus 1.1 mli
ohms um times uh and then actually you know what for fun let's do 200 amps um
just to illustrate like the flaw of this whole system that Intel came up with so
times 200 amps um you know and then that works out to 220 molts so that's going
to be like uh one .38 volts right so this is what the
voltage regulator actually aims for so 1.38 volts over here and then the board impedance is 0.5 so the vrm is going to
be at uh 0.5 * 200 that's 100 so this is
going to sit at 1.48 volts right the board impedance is
irrelevant if the impedance of the board was one uh 1 milliohm um then 1 mohm * 200 amps is uh
200 molts so instead of 1.48 we would sit at 1.58 this still doesn't match the vid
vid request right because it doesn't matter if the board was like if the
board was uh 0 milliohms right uh well 0 * 200 amps is
still Zer so 1.38 volts this is of course physically impossible because the CPU socket it's
self has more more than zero mli ohs of resistance so just just having like and
even if you were like oh well instead of a socket let's use solder balls well guess what solder balls have resistance
they're not superconductors and here's another sort of uh fun fact the
resistance of the motherboard changes based on operating temperature quite significantly because Copper's like
resistance changes with temperature a lot so if your motherboard is at uh room temperature uh the resistance resistance
of the board is actually lower than say 2 hour like 2 hours of rendering workload later right if the board
temperature goes up by like if you're yeah like if you go from 20° to 100 uh
to 100° cus the output voltage at the inductors is going to be significantly different the voltage at the CPU is not
going to change because the voltage regulator measures the voltage at the CPU and it applies the load line as a
Target right it doesn't apply the load line as like like it's not meant like
the load line slope from the voltage regulator isn't meant to compensate for the resistance of the board um that's
not what it's like it kind of it is kind of meant to do that but not in steady state right if the CPU is pulling a
constant 200 amps the low like the the impedance of the board is utterly irrelevant it doesn't matter the only
time the like the the impedance of the board uh is and the reason why we try to match
the load line to the impedance of the board is that if you have a load like if you have a impedance
mismatch right so let's say you have a um
um Z Point like let's say you have a 1 Milli ohm board uh
right and you have a 0.5 milliohm load line at the vrm
what you're going to get is uh this so you have all this
undershoot right and then on the load release you're going to get a bunch of overshoot so you get all of this and all
of this because of your impedance mismatch because the voltage regulator is like well um you know like here you
might be at 1.2 volts let's say this is 100 amps um so here you're going to be at 1.25 but the board like the impedance
of the board causes unders into the 1.1 I mean not 1.25 is going to be 1.15 but
the board is undershooting like the impedance of the board causes unders shoot into the 1.1 Volt range right so
if you have an impedance like if your load line is shallower than the board impedance you get
undershoot it doesn't affect the average voltage that the CPU sees and if you do
the opposite where like let's say you know you swap the swap the the load
lines around um or like we swap the impedance and the load line from 1 Milli
ohm and uh 0.5 we go 0.5 uh equals
board um and 1 milliohm you know equals uh
vrm LL um then all that's going to happen is you get
this like there's no like there's no extra undershoot because it's like like the
basically the voltage on its own jumps up to about here and then the voltage regulator pulls it up the rest of the
way so actually you might get something that looks a bit like your wave for might look a bit funky it might look
like this right um and actually on on this end you might get something like it'll
give you like more of a bowl shape because the voltage regulator will have to pull the volt like you you'll get the
initial 0.5 milliohms of undershoot and then the voltage regulator like has to manually pull the voltage all the way
down to hit that 1 milliohm load line and then same thing on the opposite end
where it like has to manually pull the voltage back up um but it doesn't cause
any overshoot and it doesn't cause any undershoot it just you're just it's not a problem like the point
the point I'm trying to to make here is that if you have a really good board and you set a massive load line on it uh
nothing really happens you just get a bunch of V Dro it doesn't do anything
um so the problem that Intel basically has is you know
um is that they need to offset their vid request by the expected V Dro so that
the voltage at the CPU is high enough for the CPU to function right cuz yeah if like the CPU
is trying like I want to run 5.6 GHz all core at 1.3 volts then uh I need to send
a vid request for 1.6 volts because here's the thing let's say you're and we can see this with the with the uh I5
testing that I did um man so apparently I have two
instances of this okay this is the right one um so you can actually see this when I do some prime 95 testing with
very high V Dro right so with zero V Dro um you can see how cinch and Prime 95
basically run at the same voltage Prime 95 is slightly higher voltage because uh
the temperature that it runs at is higher right Prime 95 pulls more current but um at uh 1.1 milliohms of load line
slope on both the vrm and the AC load line you'll notice that Prime 95 actually runs at less voltage than cin
bench which is a nuts because Prime 95 needs like from a overclocker like if you're static
overclocking right Prime 95 needs more voltage for stability than cinebench does but the problem that causing this
like higher operating voltage in cinch is that the CPU like we've B like the
voltage regulator has one 1.1 mil ohms of V Dro um and so you know for
cinebench all of the cores are active it's an all core workload so the CPU is sitting there and like going like well
you know what if cinebench suddenly turns into Prime 95 right because if cinebench was
running at the prime 95 voltage um then if suddenly cinebench trans
turned into Prime 95 the voltage would be too low right you would go from 1.0 N9 to like
1.06 and then you would crash at least if the CPU was bined within 30 molts of
not being stable anymore which it isn't but
um so that's like the whole concern and this is like the whole issue that Intel
has with their with their like power delivery on LGA 1700 is that they have a
load line slow like they're using a load line system that was good for like
sandybridge CPUs and LGA 1366 right where they
pulled maybe 100 amps out of the box um but with a 14900 K like you have this
very real problem of like well okay I have a 24 core like I'm running an application on 24 cores that pulls 200
amps but like there's like the slightest possibility and this isn't even that bad
right but you have this like okay it's pulling 200 amps what if it suddenly
pulls 300 right so you need that extra you need that extra voltage so that if it
does suddenly pull 300 you still have enough voltage to maintain stability
um but uh the really big problem that you run into is that you can get all of
the cores active without there being much current draw right you'll if you watch the videos where I actually catch
like the really big voltage spikes with the 125 micro code those voltage spikes happen at the start of cinebench 15 and
at the end of cinebench 15 at least that's the most common place that you see them you also sometimes see them
during Windows startup uh other cinebench versions also do it sometimes
um so yeah like various all core workloads can basically randomly unload
while keeping all of the cores still awake and when that happens the CPU is
sitting there like well the cores are active
so theoretically any microsc
I could start running Prime 95 small ffts right and if I start running Prime
95 small ffts I will pull 300 amps and
if I pull 300 amps and the voltage regulator is set to 1.1 mohms of load
line slope then the voltage is going to droop by 330 molts so I need to send a vid
request high enough that even if the voltage does does droop by 330 molts I
don't crash and so like any allore workload
because of this ends up sending the same vid request as Prime 95 actually Prime
95 eventually ends up with a higher vid request because of the higher operating temperatures but
basically uh literally any allcore workload has like the same vid request
as Prime 95 um if they were also at the same temperature which they aren't but
yeah and the most extreme examples of that like 24 cores active but not actually doing anything is funnily
enough the Asus bios yeah um the Asus bios uh for some reason in certain like
with certain settings in the Bios uh actually when it starts up it very
consistently does this like booting an Asus motherboard very consistently like if you're on a micro code older than 129
will just smash 1.6 vol into the CPU um assuming that like that that is assuming
that you set the load line to 1.1 milliohms cuz if you're like well we'll get to that um but if you have a 1.1
milliohm load line on an Asus motherboard on a micro code older than 129 it will just smash 1.6 volts into
the CPU right as the board starts and then uh sitting in the Bios it's
actually not that bad but if you I think turn off C states in the Bios or something uh then you'll actually sit in
the Bios at 1. 6 Vols because all of the cores are active
and the 14900 K's power management is going like man any minute now I could start running Prime
95 got to have that voltage but you're sitting in the Asus bios and so you're
not actually pulling any current and so you get massive voltage cuz there's no
you know like if if the CPU is pulling 10 amps there's not going to be any V Dro to pull that like vid request down
um and so the CPU gets 1.6 volts and the CPU very rapidly degrades because of that um but the other thing is like yeah
if you're running all core workloads in Windows that don't actually Hammer like don't Peak very high current draw you
get very high voltages um because this load line system is not fit for purpose with 300
amp CPUs like this is just not viable you can't do this
um so yeah like and the thing is like I understand how Intel ended up with like
this setup of like hey um we're only going to you know we're going to scale the vid request based on how many cores
are active so that if you're like single core boosting it's not actually like it's not going to request Insanity most
of the time um but the problem is like yeah but what if you get the prediction
wrong cuz you can't see into the future just because all of the cores are active doesn't mean you're actually going to
run Prime 95 small ffts next right you might just run cinch you might run something that's less heavy than cinch
right cinch still pulls like um 80% of the power that Prime 95 does so you're
going to get like 80% of the V Dro that Prime 95 does um but if you're running
something that's like 50% the load of cine bench or God forbid the Asus bios which doesn't pull any current and the
cores are all active you just get massive voltages um and so this is just a really
like clunky way of of dealing with this right um and the thing is if you have a
garbage tier motherboard like you still have this problem like this isn't a problem of like one if you put 1.1 mohms
of load line slope on a high-end board the board just has more V Dro than it needs right it's not like if you put
1.1 Mill ohms of loadline slope on a garbage motherboard that actually has 1.1 mohms of like board
impedance if you're pushing you know 10 amps into the CPU and the CPU sends a
1.6 volt vid request well it's going to get 1.58 something
volts right cuz it does like this doesn't matter so
so this is the this is the thing um is just like yeah this this load line system is like I am surprised that Intel
took so long to add a vid limiter to the micro code because the massive like 1.6
volts and higher voltage spikes that you know I measured with the oscilloscope on the 125 micro
code they're just inherent to this power delivery system like you're just going to run
into situations when all of the cores are active but nothing is actually running and when that happens you get
1.6 volts I I don't know how they thought this was a not going to
happen right this is also why they're like oh you mustn't use high performance mode because high performance mode might
keep call keep the cores active even if they're not doing anything and then you get 1.6 volts and it degrades your CPU
it's like hey Intel have you considered maybe designing a power delivery system that doesn't have the potential to wreck the
CPU if it just happens to be running in a slightly weird
situation um which apparently they are capable of doing that as demonstrated with the 129 micro code but it's just
like how did they come up with this and the thing I really really don't
understand is that in all of my testing even if you set the boards to
zero milliohm load line slope so massive undershoot right um and then you set
your AC load line to zero milliohms uh I've never seen this actually negatively affect it
stability now I wouldn't recommend that you actually do this on most motherboards because you're voltage
regulation ends up looking like this
um I'm I'm too far down the page but yeah like you you end up with massive overshoot like you end up with um like
even on an apex at level s LLC which level seven isn't zero it's like 0.2
something um at level seven there's like 100 molts of overshoot there's 100 molts of undershoot uh the CPU is still stable
cuz Intel does have some clock stretching capability uh not to the degree that Nvidia does nvidia's clock
stretching capabilities are wild like if you you can get Nvidia gpus to like run
at like 0.7 volts while reporting a operating frequency of gahz because of
how much they can clock stretch um you can't do that with Intel CPUs if the average voltage gets too low Intel CPUs
will still crash um but um like short
spikes of undershoot the Intel CPUs seem to actually handle those just fine um
but yeah even on highend boards if you set your load lines to zero uh the voltage regulation gets pretty bad and
if you had like a crappy motherboard where say the control Loop of the voltage regulator isn't super stable uh
you could end up in a scenario where like rapid transients cause the voltage regulator just to go completely out of
whack um and so yeah like I wouldn't recommend that you actually like set
super low load lines everywhere because while it doesn't destabilize the CPU it could actually destabilize the voltage
controller um which would be worse cuz uh yeah
um um you'd be back to basically degrading this CPU very rapidly um if
that happened um so yeah so this is like
like and this is probably like you might remember I made a video about how like Minecraft servers are burning through i
94900 k like super fast right and in that video we actually looked at like
the hardware info report from those super micro boards well those super micro boards use a 1.1 milliohm load
line slope and they're on really old micro
codes so yeah those boards like I don't like that if that Minecraft server software keeps the cores
awake um hello 1.6
volts like and the fact that like like so yeah
this is just as far as I'm concerned this is just an inherently unsafe power delivery system and Intel with one like
with the 129 micro code like finally decided to re it in right it's like you
know what maybe we shouldn't give the CPU the ability to kill itself just just
an idea um now the other thing that was going on
with u LGA 1700 um which
uh is like possibly why it took so long for like people to notice this is the AC
load line undervolting um so this is a separate issue this is why like you may remember
there was like articles about uh like people Mass returning I9 in Korea when
Tekken launched cuz yeah uh I 9's like you couldn't get through the Shader
compilation for that game with an I9 and the reason why that happened is uh so
let's say uh Asus um so Asus for example uh had the
following uh vrm settings vrm
LL equals 0 point uh no actually equals 1.1
milliohms um AC LL equals
0.55 and then DCL equals 1.1 now DCL is literally just
used for calibrating power measurements so it doesn't affect the uh vid requests
at all so that's why DCL like in the Asus configuration it's 1.1 to match the
vrm because if DCL like if we set this to four um this will make the CPU think
that it's pulling way less power and if we set this to zero um it'll make the
CPU think that it's pulling more power so if this isn't properly matched to the vrm you basically it's like add like
it's basically like changing the power limit is what the DC load line does uh if it's not properly matched to the load
line of the vrm um so that's like so basically the DCL
is generally well matched now gigabyte which I'm just going to go for GB would
have a vrm uh LL of 1.1
at least on the boards that I've tested it might be different on the ON Semiconductor boards uh an
ACL of 0.4 milliohms um and a
DCL of uh I think 0.9
milliohms so for some reason gigabyte is actually over reporting the power consumption to the CPU a little bit um I
kind of suspect that gigabyte just doesn't know what these two do so so that's how they ended up with
that um and then MSI now MSI unfortunately I have no idea what their vrm LL is um so
vrm LL equals no idea um I haven't like
I haven't looked into figuring out what it is yet acll equals 0.5 and
DCL equal 0 .8 um so I'm guessing the vrm is probably somewhere close to 0.8
because I suspect MSI it is a bit better informed about what DCL does than
gigabyte is um but yeah so this this is probably
like 0.9 or 0.8 um or yeah like it it might be
something like that um but anyway so the result of doing things like this with your loadline settings I'm going to run
out of space so we're just going to scrap MSI cuz it doesn't matter um we're just going to go with like Asus for the
the calculations so the reason that even a brand spanking new I9 Factory sealed
you know you have like the the little sticker Factory steal still there brand new I9 CPU you stick it into your
high-end Asus motherboard and then you try to run Tekken and it can't compile shaders well the reason that happens is
because the CPU um would make vid requests with the
assumption that there's 0.55 milliohms of vrou but actually there is 1.1
milliohms of eou so what would happen is um let's say the vrm uh or let's say the
CPU is pulling
uh so let's say we're running we're trying to run 5.6 all core again right so
5.6 um just for consistency and we're going to do it at
uh 1.3 volts right so at zero amps right
we just request 1.3 volts because there's no V Dro um but at
a 100 amps right to get 1.3 volts um we do
okay 100 amps time 0 why did I go with Asus let's just take
that off um uh 0.5 so we need Plus 50 m
Vols now obviously I'm I'm leaving out the part where the actual current that is used for the AC load line isn't the
current that the CPU is currently pulling but the current that the CPU thinks that it could be pulling but for
like for Simplicity sake we're just not going to worry about that right now right so CPU is like okay I'm pulling
100 amps uh so I need to you know there's 1.5 milliohms of load line slope
uh I'm going to need Plus 50 MTS so it sends a vid request for 1.35 volts the
vrm gets that request right vrm gets this 1.35 volt like so
this is the svid um so it sends 1.35 volts over svid
um the vrm gets that goes minus 1.1 milliohms uh time 100
amps and you end up with uh 1.2 [Music]
four yeah you end up with 1.24 volts so you know CPU needs 1.3 volts to
function actually gets 1.24 but this is not a big deal a 60 molt undervolt is
probably not going to crash your I9 um but let's say the CPU is pulling
200 amps
uh so that's going to be you know uh plus 100
molts so we send a 1.4 volt esid request right um and i' I've really
botched this um so we send 1.4 volts over
Sid um and then we do minus 1.1
milliohms time 200 amps
and well that's 220 MTS so we end up with uh 1.4 minus 20000 is 1.2 minus 20
we end up with 1.18 volts
ah well this is like 120 molts worth of undervolting the CPU might actually
crash at this point um and obviously as the current draw increases right this this trend
continues of like your you basically like basically what this does is your CPU gets more and more undervolted the
more power the CPU is pulling so if you're just doing stuff on the desktop
you don't really have an undervolt if you're running mild workloads you don't really have an undervolt and then when
you hit something really really heavy suddenly the CPU is getting like 120 MTS
less than it should be or maybe even 200 molts less than it should be cuz yeah let's let's do the math for 300
amps um right so this is going to be plus
150 so our request is going to be 1.45 and then uh it's going to be
negatively offset by 330 M volts which is going to be uh
1.0 cuz minus like it's going to be minus 1.1
I'm just going to write it out I don't this doesn't actually help me do it in my head does it but whatever um so
that's going to be 1.1 1.15 uh so we're going to end up at 1.1 uh 2
volts right so you get this like progressively increasing undervolt the
more power the CPU pulls and so if yeah there's a big current Spike for a
fraction of a second while running some workload um which you know when the CPU
like this is a a funny thing is you're going to like that five like Intel's power management is really fast um so
for comparison like the shortest uh Power State transition I've seen on like AMD CPUs is a millisecond so like the
voltage level will change every millisecond based on temperatures power consumption that kind of thing Intel as
far as I can tell can do it every 0.1 milliseconds so you could actually end
up in a CP situation where the CPU boosts to 5.6 GHz for like 0.1
milliseconds blips 300 amps gets a big fat undervolt and
crashes lovely um um right so that's kind of the and so
like when you're running other workloads like say cinebench 23 3 which runs at like 5.2 GHz it doesn't look like
there's anything wrong with the CPU but if you run something where you know the workload is very spiky uh it's suddenly
really unstable because for brief moments of time the CPU is getting like
depending on how like how much Peak current there is you get different under volts
um so yeah not great so this is like the AC
load line undervolting was happening now the funny thing about these AC load line under volts is that they do kind of uh
prevent the problems that Intel's loadline slope system has right so the
the gaming motherboards are actually less susceptible to the or at least used to be less susceptible to the really
high voltage spikes until the 125 micro code came out because they were undervolting the AC load line right like
this whole issue of like oh if 24 cores are active 307 amps time 1.1
mohms well it's a lot less of a problem if you just
0.55 right and so this is now like plus 150 uh Mill volts instead of plus 300 so
now you're like 1.3 volts plus 150 is like yeah not not uh you know you don't
get like you don't get the massive like the voltage spikes are somewhat reduced there are still there um like some of
the like well I mean 12 the 125 micro code is like especially bad um that one
as far as I can tell is probably actually the least safe cuz Intel basically went and like they blocked the
AC load line under VTS um but they didn't like you don't have the vid limiting yet so you're just left at the
mercy of Intel's just ridiculous loadline slope system um
but on on the but even on the earlier micro codes you do still get some pretty
pretty nasty voltage spikes um they're just not quite as bad as they are on the
125 but then you have like OEM motherboards right like the super micro
servers those don't do any AC loadline undervolting right they run an 1.1
milliohm AC load line I'm sure if I hooked up the if you hooked up the a to one of those super micro boards it would
be hitting 1.6 volts all over the place
um so and this does kind of make me wonder how bad uh this is for like
system integrators like HP Dell Lenovo cuz the motherboards they build are
trash and I would not be surprised if every single one of those boards had a 1.1 mohm loadline slope so that they can
shave off as many capacitors as physically possible um but if those boards have a
1.1 mohm load line slope that means anytime the CPU is has 24 cores active uh it's
liable to request like 1.6 volts unless they're manually lowering the ICC Max in
the in the Bios um so yeah like Intel's and also what's
crazy is because the AC load line uh like so as I with the AC load line
undervolting right like Asus gigabyte MSI they were all getting away with like
100 molts plus under volts right out of the box right so Intel is shipping these CPUs with a big fat safety margin from
the factory right like your I9 can probably undervolt by 100 molts if it's
using the like Intel AC load line like AC load line
recommendations um so if you have a like crappy OEM
motherboard with 1.1 Milli ohms of load line slope that CPU actually is running a bunch of extra voltage most of the
time right because anytime the CPU isn't pulling the expected Peak current it
actually has a bunch of sa like more safety margin than it normally does so if the CPU is degra like the point at
which you notice that the CPU has degraded is when all of the safety margin is gone right so if the CPU comes
from the factory with 100 molts of safety margin and it loses 5 molts of
safety margin every two months it'll take like uh 50 months before you notice that
the chip is dead um and that's assuming that it doesn't have a bunch of extra voltage due to the AC load line right
cuz like like basically your your safety margin is going to like wear away uh
slow for something like cinch than it is going to wear away for something like Prime 95 cuz when you hit Prime
95 um you end up in that um I'm going to bring back
the peak current thing hopefully I grabbed the right one uh okay I did um because I have two
instances of this open for some reason um so like here right like this is the
same loadline slope if you if you always just stress tested the CPU with cinebench
um cinebench would actually tolerate 30 m volts more degradation than Prime 95
does right so if you never load the CPU to 100% you're it's going to take significantly longer to notice that it
degraded and so yeah like this is
like like this situation is just kind of wild because you have so many different
problems at the same time like for one thing you have the AC load line under volts from the motherboard manufacturers
which destabilize even brand new CPUs with no degradation then you have the fact that
Intel's loadline like implementation for LGA 1700 is actually insane and
inherently unsafe so if you put it under the wrong kind of workload it'll just try to kill the CPU um and then because
this like load line system also just like overv volts the CPU which is like yeah that's that's just what it
does um if you have a like it actually delay like it the more load line you run
the longer it'll take for you to notice the degradation caused by the excessive
voltages like and it took them two years to
realize that maybe maybe it would be a good idea to prevent the CPU from sending vid requests greater than 1.55
volts like I actually just
cannot so the main thing I want to make clear is that the degradation of these
CPUs is well okay I'm I don't know like I don't know for sure that like CU
other things the motherboard of vendors were doing was disabling things like C which limits the peak current that the
CPU can pull um they were disabling ICC Max they were disabling uh the power
limits though I really think the current limits are more important than the power limits but whatever um so they were
disabling the power limit so you know and the thing is there's multiple types of degradation that you can suffer in a
in a chip if you just heat a chip to a high enough temperature at some point you'll kill it with temperature alone
but you can also destroy a chip with just high voltage right which is what Intel's load line implementation will do
if like the the thing is like that 1.1 mohm number is literally in Intel's own
documentation as like the maximum load line that you can use and it's just like
but if you use that number you get 1.6 volts
like what are they doing how how did that number make it into the documentation what's more crazy is that
for like laptop CPUs uh that load line Max load line slope is like 1.7 or
higher even so like and then they have this like
predictive vrou compensation where it's like yeah and if the prediction is wrong then the CPU dies amazing wonderful just
truly I don't know what they're doing um I mean I understand why it works the way
it does CU yeah it it's kind of awkward if your CPU like if the motherboard has a 2 milliohm load line slope or a one
one mli load line slope and then it vrops to the point that the CPU crashes yeah that's probably not ideal so yeah
you do kind of need to make sure that the CPU has enough voltage but maybe don't send vid requests that will kill
the CPU if the vrou doesn't happen um just saying uh you know that might be a
good idea uh anyway where was I going with this right
um that was a reflexive right um yeah so you can have like multiple
different forms of degradation right you have degrad like you have degradation due to excessive current degradation due
to excessive voltage degradation due to excessive temperatures and the motherboard vendors were disabling
current limits and power limits so that would like if these CPUs were also susceptible to excessive current then
disabling things like ICC Max would cause the CPUs to degrade when running all core workloads at even very low
voltages right because a 14900 k um you can pull that 307 amps uh at like one
Volt or okay not one volt but like 1.1 volts if you run something like Prime 95 small ffts like yeah you can pull 300
amps um so you don't need a crazy high voltage to pull insane amounts of current on a 14900 K so if the 14900 k
is susceptible to electromigration degradation at like 400 amps or something then yeah turning off the
current limit is probably not good for the longevity of the CPU um but at the same time the CPU is definitely
susceptible to degradation due to excessive voltages and that is entirely Intel's fault because this VR
compensation system is like like why would you bolt this on top
of like like this is like this is why Nvidia doesn't do this right like nid
like if you think about a 4090 right a 4090 probably has a peak current
somewhere in like the 500 600 amp range Nvidia completely gave up on this style
of load line system with the GTX 7 uh 900
series The 980 TI has a like better power delivery
system than a 14900 K because with the because gpus are
inherently just extremely noisy um in terms of like transient loading so yeah
if you used an Intel style freaking load line system on a GPU like you would never be able to make a functional
GPU um or at least not a fast functional GPU you could always clock it so low
that you know it wouldn't need very high currents or something but um yeah like if you if you applied this
style of power delivery to a 4090 like it would just it would just not work at all um so and on that note AMD also uses
a more Nvidia style power delivery system these days where the motherboards come with extremely shallow loadline
slopes and then the CPU just clock stretches whenever there's undershoot um
but Intel has this like Sandy Bridge era loadline slope system and then they're
like well now we have 300 amp CPU so I guess we just compensate for 300 amps worth of
VR with at 1.1 Milli ohms that that seems totally sensible um and if the vrou doesn't
happen then I guess the CPU just dies
um so yeah like um so like this like the voltages
the insane voltages Intel's fault uh the no power limits maybe the motherboard
vendor fault personally I don't think that the current limits and the power
limits have anything to do with most of the degradation that people are seeing because again those super micros servers
had all of the Power limits all of the current limits the only thing they did wrong was they in they followed Intel's
load line uh load line specs that was the only thing that you could fault
those super micro boards for is like yeah they they like followed the specifications that Intel put put into
their documentation is like so that's Intel's fault
um and uh is there anything
else so um yeah but then the other point is
that a lot of the like out of box instability on brand new CPUs is actually just the motherboard vendors undervolting the
CPUs um so that's like a separate issue that's not actually degradation if
anything the fact that like Asus gigabyte and MSI and they were all undervolting the CPUs probably slowed
down the degradation significantly on the gaming motherboards so it like admittedly you
know they turned off all the power limits and stuff but at least your CPU was getting undervolted a bit um they
also removed things like uh TVB uh was like TVB is supposed to prevent your CPU
from boosting to 6 gigz if it's above 70° celius yeah that didn't work on like
some past micro code versions so the CPU would still go to 6 GHz even at like
100° um so you
know but like the high voltage spikes that's Intel that's entirely Intel
because no motherboard that I've tested tested is like adding extra voltage to
the output that's not a thing
um so yeah and I really don't get why
Intel didn't just like they have clock stretching in these CPUs as far as I can
tell why didn't they just recommend like why didn't the spec just say that you're supposed to use like a
0.3 load line slope for everything
cuz like they they have to like the thing is I don't even know for sure what the like expected Peak current that is
used for the vrou compensation is like again with the the I5 data that I have it seems to be more than the ICC Max
which would make sense um right like the ICC Max for an i5 at like
stock is supposed to be 200 amps as far as I know but you can see that the peak voltage actually goes up all the way to 250
it doesn't go up at 500 amps cuz evidently that's not you know there's no
way an i5 is ever going to pull that much current um so it doesn't need to compensate for that but I guess it like
could pull 250 or 225 and so it does compensate for that much
um right and you can see how like as the ICC Max get slower the vid requests get lower
um though some of that is also just going to be like power limiting like the CPU getting power limited so like not
reaching the higher boost States for some of this um but
uh yeah so like like my point is like I don't even know what the the the peak
current that Intel is doing the compensation with is Intel knows what
that is like they designed this and if you do the math on it and go like hey
but what if what if just freak scenario right for some reason some software lights up all of the cores but doesn't
do anything with them what happens to the CPU in that scenario is apparently a question Intel never
asked and also that scenario happens like anytime you start or stop an allcore workload which is just kind of
funny like so it's actually kind of common and it's just like yeah you know
it'll be fine we just you know um so anyway that's there like now now I've
gone gone through everything uh I think and uh yeah
like key takeaways is like this the voltage regulator measures the voltage
at the CPU okay like that that's one of the other like big frustrations I've had
with seeing a lot of the discussion is like the the the load line is like
the the impedance of the board in steady state is irrelevant the voltage regulator will compensate for that on
its own but it can't compensate for it during fast load Transitions and that's
why we have load line slopes um and then because Intel for some reason went with
really deep load lines probably because the like here's the thing the socket itself probably has like 0.2 milliohms
of resistance right so if you if if um if you set like oh 0.2 milliohms as
your uh load line slope like your voltage regulator would basically have to live on the other side of the CPU
socket right like where the back plate is that's where the voltage regulator would have to be and even that probably
wouldn't be enough cuz again the voltage regulator itself has lag which basically translates into more
impedance um so
yeah so this is just like there um anyway but yeah like my main
frustration is I've seen people saying like oh the load line slope is for compensating for like the like that's
not what it's for it's for load transitions if your CPU pulled 300 amps
all day every day we wouldn't need load line like we literally wouldn't need load lines but because your CPU doesn't
pull 300 amps all day every day uh we do need load lines because when it goes from pulling very little current to
pulling a lot of current uh you need to you know
like deal with that um and load lines is how uh well we've been dealing with that
for several years uh unless you do what Nvidia and AMD did where you just let
the chip deal with the fact that the voltage is really low for a bit um
anyway uh there okay now I think I've actually said everything I wanted to say uh so that's it I guess um so thank
you for watching hopefully at this point it makes sense um like what is going on with load
lines on LGA 1700
um and uh yeah man I don't know why I'm
struggling so hard with the outro here I'm just worried that there's like some little that I might have missed and
I'm going to get to read some incredibly stupid comment after recording an hour 40
minute video and somebody still isn't going to get
it

Click to expand...

coercitiv said:

Oh Buildzoid, why do you have to do this? Another 1h40m video I HAVE to watch because I want to know about this.

I guess no movie tonight for me

Click to expand...

DAPUNISHER said:

It's good to see Starforge has extended the warranty. Though that should be expected with how big their markup is.

Click to expand...

Following the recent warranty extension announcement for affected Intel Core 13th and 14th Gen desktop processors, Intel confirms these currently available processors are not affected by the Vmin Shift Instability issue:

• 12th Gen Intel Core desktop and mobile processors
• Intel Core 13th and 14th Gen i5 (non-K) & i3 desktop processors
• Intel Core 13th and 14th Gen mobile processors – including HX-series processors.
• Intel Xeon processors – including server and workstation processors.
• Intel Core Ultra (Series 1) processors

Click to expand...