Question AMD Ryzen 3950X boost behavior

[Kocicak]

My 3950X I got today seems to have a weaker chiplet in it or I have no idea what is happening. First 8 cores routinelly boost over 45x, the second 8 never hit 44. Did I get a lemon (again)? I am attaching HWmonitor screenshot after few minutes of running light load.

BTW after a few minutes of running after I screenshotted it, 2 of the second 8 cores hit 4341, they are not all the same as you can see in the picture.

 

[amrnuke]

3900X was not hitting the advertised boost frequency at all. HALF of this 3950X works as advertised and as I expected, the other half surprised me. 300 MHz is quite a notable difference.

Sorry, but your expectations are wrong. The entire chip works exactly as advertised. The chip is designed to run at 105W, with a 3.5GHz base and a 4.7GHz boost. And it does.

Let me ask you this - for the 3800X, for a large # of people, not all cores on the chiplet can boost to 4.5 GHz. However, it hits all the advertised ratings. Do you see this as a problem?

 

[rainy]

HALF of this 3950X works as advertised and as I expected, the other half surprised me. 300 MHz is quite a notable difference.

This is simply not true:

Max boost for AMD Ryzen processors is the maximum frequency achievable by a single core on the processor running a bursty single-threaded workload. Max boost will vary based on several factors, including, but not limited to: thermal paste; system cooling; motherboard design and BIOS; the latest AMD chipset driver; and the latest OS updates.

https://www.amd.com/en/products/cpu/amd-ryzen-9-3950x

 

[UsandThem]

3900X was not hitting the advertised boost frequency at all. HALF of this 3950X works as advertised and as I expected, the other half surprised me. 300 MHz is quite a notable difference.

Not, it is not.

You will never notice some of your cores running 300mhz lower (in micro-seconds) than what the best cores boost to (especially when all you seem to do with your computers is to install an HEDT CPU and then come to the forums to tell us how it's a "dud").

It gets really old, really fast.............

Maybe you should return the 3950X and buy a i9-9900KS instead? Maybe the all-core 5 Ghz boost speeds will satisfy your expectations? Even better, buy a pre-binned 9900KS from Silicon Lottery, and boost away.

 

[Topweasel]

3900X was not hitting the advertised boost frequency at all. HALF of this 3950X works as advertised and as I expected, the other half surprised me. 300 MHz is quite a notable difference.

Do you get 4.7GHz on any cores?
Do you get the 3.5GHz on all of your cores?
Do you get within a little wide margin of error of other peoples performance?

Did you use recommended cooling?

You are not on the last one so anything else quite literally has to be ignored. Now they recommend water cooling because making sure you get the right air cooler is a lot harder then saying use a 280mm AIO, though I am sure a 240mm would work (my 3900x with a H100i seems to do well enough).

AMD never promised and after the ordeal with the 3900x you knew better then to expect to get anything but a handful of cores to hit max boost. It's a single core boost and may only be available on one core at all. You said you have seen it on two, so lucky you. Everything is within specs. You didn't get a dud. Now read up on Anandtechs, GN, and Level1's, articles regarding the boost. Decide if you want a real beast of a CPU, or going to continue to make waves about something you already got slapped down about with the 3900x, if you are you might as well return it and shut up. You knew what you were getting or at least had no reason not to know and are nearing official annoyance territory and can join the full other 2 people I have put on ignore in my nearly 20 years on this site.

 

[BigDaveX]

Okay, I was willing to defend the people criticizing the 3900X failing to achieve its advertised boost speeds, since it wasn't delivering what was promised, and I know for a fact that Intel would have been flamed to a crisp had their chips been displaying the same behaviour.

But this time, AMD has nothing to answer for. They advertised the CPU as being able to clock up to a certain speed when the load was one or two threads. That's it. They didn't advertise all the cores as being able to achieve that clockspeed, they said at least one of the CPU's 16 cores could achieve 4.7GHz under light load conditions.

Honestly, this is like complaining about... actually, I'm not even going to bother with an analogy, since it seems these days that whenever you accompany an argument with an analogy, people just completely ignore the underlying argument, nitpick the analogy, and then think they've also disproved the argument.

 

[Topweasel]

Okay, I was willing to defend the people criticizing the 3900X failing to achieve its advertised boost speeds, since it wasn't delivering what was promised, and I know for a fact that Intel would have been flamed to a crisp had their chips been displaying the same behaviour.

But this time, AMD has nothing to answer for. They advertised the CPU as being able to clock up to a certain speed when the load was one or two threads. That's it. They didn't advertise all the cores as being able to achieve that clockspeed, they said at least one of the CPU's 16 cores could achieve 4.7GHz under light load conditions.

Honestly, this is like complaining about... actually, I'm not even going to bother with an analogy, since it seems these days that whenever you accompany an argument with an analogy, people just completely ignore the underlying argument, nitpick the analogy, and then think they've also disproved the argument.

I'll admit to being one of those at 4.5 GHz+ whats the last 50MHz or 25MHz really worth type guys. But that was a legit issue that got fixed. I was just telling people I was happy with my perf. But his issue before went beyond that. But even if that was his issue before, I don't know why he spent more going up and getting something that is going to have a harder time sustaining boosts. It all gets thrown out when he says he hits his boost target. Pure trash after that.

 

[Zucker2k]

the other half surprised me.

 

[Hans Gruber]

Benchmarks mean everything. The 3950x is a monster chip. I can't believe people would complain even @ 4.5ghz with a 3950X.

 

[therealmongo]

Some people can't be helped.

If your being serious then you need some professional help as it seems you are not able to distinguish between reality and your own ego centered reality.

It's very simple, generally when we want to purchase something we find out about the item so that when we buy it we know what the manufacturer has sold us.

You are completely neglecting this irrefutable fact either through sheer ignorance, below average intelligence, psychological disorders or you are trolling.

What's even more depressing is that its the same characters who are 'liking' your posts who are unfortunately assisting you on insisting on your self made dissolution.....

 

[DrMrLordX]

Ok, when I force one R15 thread on the second chiplet, the maximal frequency is 4391MHz (before that it was 4341).

1T R15 score on the first chiplet is 212 and on the second 203.

Okay, that's pretty normal, and slightly better behavior than what I get with my heavily watercooled 3900x. My "bad" CCD hits maybe 4.3 GHz in a ST workload (SuperPi mod 1.5 xs). I haven't tested it with CBR15 but I suspect it'll behave similarly. I might be able to get the "bad" CCD to boost higher with PBO but it's not worth the effort.

Now the thing to understand is: you won't ever get hurt by this. The "slow" CCD is still fast enough in default operation to maintain all-core boost frequencies (which for the 3950X are around 3900-4000 MHz, depending on cooling, in stuff like Blender or what have you). Test it out if you don't believe me. The scheduler is smart enough to move lightly-threaded stuff to the "fast" CCD so you won't ever notice performance slowdown from ST workload threads migrating to the "slow" CCD. AMD knows this, and uses less-adequate CCDs to fill out the 16c for the chip to aid in the binning process. Pretty clever of them actually.

AMD Disco Cooler is sometimes called Wraith Prism and it does not come with 3950X. I bought it extra. 3950X comes with FOAM.

That's pretty good actually. I figured Wraith Prism could handle a 3950x. Now I know it can. You would have been better off with an NH-D15 though.

When running light load, CPU voltage reaches 1.50V

Also normal for low-current workloads. What's the voltage in CBR20 MT?

The CPU adjusts voltage and frequency according to the load, now it runs Prime95 at 0.919V and 3200 MHz, consuming 120W at temperature 68°C. The cooler is pretty silent too.

Interesting! What's the clockspeed/voltage in CBR20 MT and Blender Benchmark? Also, if you want more MT performance, I can probably help you out without forcing you out of default mode. No PBO either. Should get you lower temps as well.

 

[Kocicak]

The scheduler is smart enough to move lightly-threaded stuff to the "fast" CCD so you won't ever notice performance slowdown from ST workload threads migrating to the "slow" CCD. AMD knows this, and uses less-adequate CCDs to fill out the 16c for the chip to aid in the binning process. Pretty clever of them actually.

It not only prefers the right CCD, it also loads the same cores all the time, one can start to wonder, if it would not be better to load other cores too not to "wear out" the best ones.

It is clever of them for them, but using a "filler CCD" in the most expensive processor of one platform seems to me from a consumer perspective as less desirable...

You would have been better off with an NH-D15 though.

I do not like to hang 1 kilo weights on motherboards if possible.

What's the clockspeed/voltage in CBR20 MT

3800-3850 MHz, 1.075V, 135W, 72°C, score a little bit over 9000, cooler set to be silent, with more aggresive setting everything would improve a bit probably (If I remember correctly, it was 68°C with faster fan settings but I did not note any performance numbers yesterday).

I may try some all in one water cooler again, but for now the air cooler is enough.

 

[UsandThem]

It not only prefers the right CCD, it also loads the same cores all the time, one can start to wonder, if it would not be better to load other cores too not to "wear out" the best ones.

You're the only user who comes to mind.

It is clever of them for them, but using a "filler CCD" in the most expensive processor of one platform seems to me from a consumer perspective as less desirable...

The same thing you said about the 3900X, so............

I do not like to hang 1 kilo weights on motherboards if possible.

Not even touching this with a 10 foot pole.

3800-3850 MHz, 1.075V, 135W, 72°C, score a little bit over 9000, cooler set to be silent, with more aggresive setting everything would improve a bit probably.

You posted this thread when you have it set to this power profile? Holy sweet baby Jesus.

I'm done. I just can't take anymore of this..........

 

[kawi6rr]

3900X was not hitting the advertised boost frequency at all. HALF of this 3950X works as advertised and as I expected, the other half surprised me. 300 MHz is quite a notable difference.

Do AMD a favor and buy Intel I'm sure they'll be fine with that.

 

[Ottonomous]

Next week on Kocicak's AMD kackles: Why one of the 8 chiplets on his 3990X dropped the boost frequency for two seconds during prime95, running on the AMD wraith cooler

 

[Ottonomous]

Do AMD a favor and buy Intel I'm sure they'll be fine with that.

This Kocicak. Buy a 10980XE since you paid 950 bucks and put a stock intel cooler on it, then come back and post if it hits the advertised boost frequencies

 

[Campy]

Buy Intel, they need all the customers they can get.

 

[Charlie22911]

I have to say that I am disappointed at the way the community is reacting here. I’ll leave it at that.

OP - The chip is performing as designed, these days the base clock is really all you are guaranteed to get. I don’t like how noncommittal manufacturers seem to be these days when it come to clock speed, but this is what we have to work with. The 9900ks is really the only chip I’m immediately aware of that advertises an all-core boost, and it’s just that; a boost.

 

[Topweasel]

I have to say that I am disappointed at the way the community is reacting here. I’ll leave it at that.

OP - The chip is performing as designed, these days the base clock is really all you are guaranteed to get. I don’t like how noncommittal manufacturers seem to be these days when it come to clock speed, but this is what we have to work with. The 9900ks is really the only chip I’m immediately aware of that advertises an all-core boost, and it’s just that; a boost.

If you didn't read the last few threads he opened then you can't really criticize. When he created the 3900x thread people were a mostly more open about it. But the gem is the 3600x2 thread he created. Read that and you will understand why most people on this forum are kind of dismissive of him. There is a distinct trend in his posting and imho I question if even got either CPU and instead parroting things about the cpu he thinks he can negatively point out. Using an AMD cooler that you can't buy retail on a CPU that doesn't ship with a cooler is really straw that breaks the camels back there. You don't ebay one when there are better coolers for less (and never really mentioned how he got his).

 

[DrMrLordX]

It not only prefers the right CCD, it also loads the same cores all the time, one can start to wonder, if it would not be better to load other cores too not to "wear out" the best ones.

It won't. That CPU @ stock will probably last ten years or longer. There were a few "suicide snowflake" 3600s out there that died fast (allegedly), but I suspect that was more a firmware issue than anything else.

It is clever of them for them, but using a "filler CCD" in the most expensive processor of one platform seems to me from a consumer perspective as less desirable...

No, it's clever for us too. You would not benefit from have a second "fast" CCD unless you plan on all-core static OC. The way the boost algorithm works, clocks will drop as more cores are engaged, to the point that you won't be boosting anywhere near even 4.4 GHz once you start using the second CCD. Without using "bad" CCDs (which, mind you, are still better bins than the 3600, 3600x, and 3700x), there would be no 3950X CPUs at all.

I do not like to hang 1 kilo weights on motherboards if possible.

I do. I've been doing so since 2009, and all the motherboards I've tried since then have handled the weight perfectly. I've had either an NH-D14 or NH-D15 on:

MSI 790FX GD70
Asus A88x Pro
ASRock x370 Taichi

They all took it like a champ with no signs of board cracking or socket fatigue.

3800-3850 MHz, 1.075V, 135W, 72°C, score a little bit over 9000, cooler set to be silent, with more aggresive setting everything would improve a bit probably (If I remember correctly, it was 68°C with faster fan settings but I did not note any performance numbers yesterday).

You should go for the AMD Performance profile and start tweaking LLC settings. In fact, it's simple: turn off LLC (or set it to its lowest setting if you can't turn it off completely) and maybe apply a small negative voltage offset. Your MT clocks and performance should go up while your ST clocks may go down by about 25-50 MHz.

 

[amrnuke]

It won't. That CPU @ stock will probably last ten years or longer. There were a few "suicide snowflake" 3600s out there that died fast (allegedly), but I suspect that was more a firmware issue than anything else.



No, it's clever for us too. You would not benefit from have a second "fast" CCD unless you plan on all-core static OC. The way the boost algorithm works, clocks will drop as more cores are engaged, to the point that you won't be boosting anywhere near even 4.4 GHz once you start using the second CCD. Without using "bad" CCDs (which, mind you, are still better bins than the 3600, 3600x, and 3700x), there would be no 3950X CPUs at all.



I do. I've been doing so since 2009, and all the motherboards I've tried since then have handled the weight perfectly. I've had either an NH-D14 or NH-D15 on:

MSI 790FX GD70
Asus A88x Pro
ASRock x370 Taichi

They all took it like a champ with no signs of board cracking or socket fatigue.



You should go for the AMD Performance profile and start tweaking LLC settings. In fact, it's simple: turn off LLC (or set it to its lowest setting if you can't turn it off completely) and maybe apply a small negative voltage offset. Your MT clocks and performance should go up while your ST clocks may go down by about 25-50 MHz.

This is very informative, and I hope @Kocicak actually reads this.

The chip is performing precisely as designed, and he is matching review benchmarks even with a borderline air cooler AND having fans set to silent mode!

If anything he actually got a gem of a chip.

@Kocicak, again, the reason your second chiplet would not benefit from being a "good" chiplet (in your words) is because of power loading and thermal throttling. As the chip adds more threads to an activity it decreases the boosts. By the time you get to the 9th core, the boost is already going to be lower than the peak boost of the not "good" chiplet, therefore having a better-binned chiplet wouldn't matter. You saw this in your testing. At MT CB20, "3800-3850 MHz". Exactly. So even if the second chiplet could boost to 4.7 GHz on its own, it never would in a real world scenario because of the way boost clocks throttle once you add more cores/threads.

 

[RetroZombie]

Kocicak, let’s check out other cpu, from the time vault:

From anandtech: The Intel Core i7-8086K Review (Conclusions: Save Your Money) our own Dr. Ian Cutness on June 11, 2018

On paper, the change in specifications are a little underwhelming to be honest. This means that the only real performance benefit users will see is when the CPU is under a single-core stress. Given the nature of PCs having multiple applications open at once or running in the background, a truely isolated single-core load almost never happens: in fact with our processor we only able to trigger a core to 5.0 GHz unless we set the affinity to a single core. In that respect, the Core i7-8086K is very limited, especially when it commands a premium price.

From Tomshardware: The Core i7-8086K Review (Rare moments of 5 GHz operation) By Paul Alcorn June 27, 2018

Intel also increased the single-core clock rate to 5 GHz. We were able to sustain 5 GHz in tasks confined to a single core, such as Cinebench and LAME. However, the busy scheduling environment in a modern desktop operating system, which finds threads migrating frequently between cores, prevented 5 GHz operation in even mainstream tests like our gaming benchmarks. In other words, don't expect to see 5 GHz very often.

From Pcmag: Intel Core i7-8086K Limited Edition By Andrew Leibman October 2, 2018

For strict single-core workloads, the Core i7-8700K can increase the speed of that core to 4.7GHz, while one active core on the Core i7-8086K will surge to 5GHz. That's undeniably an impressive number, but in 2018, single-core workloads are becoming increasingly antiquated.

From techradar: Intel Core i7-8086K review By Kevin Lee July 20, 2018
5.0GHz boost clock refers to single-core Turbo Boost, Benchmark scores equal to Core i7-8700K

In our testing, the 8086K performs no better, and sometimes even scores a little lower, than the 8700K in every single benchmark. We were so baffled that we reran every test five-times over just to make sure it wasn’t a fluke, but sure enough this 5GHz processor never pulls ahead of the chip it’s designed to outpace.
Meanwhile, the rated 5.0GHz only refers to the chip’s single-core Turbo Boost, which is a capability few, if any at all, programs will take advantage of, unless they’ve been coded specifically to only push a single core. Cinebench R15 has one of these such processes, but as our testing shows, the 8086K scores a point lower in a benchmark it should have aced.


Resuming:
Do you have a smartphone? It’s normal to not see the clocks hit the specified by de cpu (or phone) manufacturer. All the cpus in mobile phones don’t guaranteed the clocks they advertize and nobody checks if they do, did you or anyone here with a smartphone even bothered?

And from the amd website: Base Clock 3.5GHz
You said: “300 MHz is quite a notable difference.”

My analysis:
Seams everything is alright, it’s running 900 Mhz above spec.

To others:
I’m surprised with the patience of some you guys here, having to reply someone that creates a new thread that is questioning something, and then came here explaining the why it is that way and that everything is running alright, and then have that same person (or even others) incorrectly arguing with the all the correctness explanation it’s actually to be applausive!

 

[lightmanek]

I don't think AMD is going to bin two high clock capable chiplets in one CPU and this is by design!

We should take into account that the chips willing to run at higher freq. are also more leaky on the transistor level. What I see with twin chiplet CPUs is one chiplet binned for clock and one for best power efficiency at high core loads. This allows AMD to hit TDP targets and also offering best possible boost clocks in a single package.

For sure there is a place for 'special' Ryzen 9 3955 TWKR edition, but that I guess would come with no warranty and 280W TDP (exaggerated).

Just my opinion formed from reading reviews and applying some logic.

 

[jpiniero]

Kocicak, let’s check out other cpu, from the time vault:

From anandtech: The Intel Core i7-8086K Review (Conclusions: Save Your Money) our own Dr. Ian Cutness on June 11, 2018

On paper, the change in specifications are a little underwhelming to be honest. This means that the only real performance benefit users will see is when the CPU is under a single-core stress. Given the nature of PCs having multiple applications open at once or running in the background, a truely isolated single-core load almost never happens: in fact with our processor we only able to trigger a core to 5.0 GHz unless we set the affinity to a single core. In that respect, the Core i7-8086K is very limited, especially when it commands a premium price.

You have to remember that on Intel you are limited to whatever they put as the multiplier limit per number of active cores unless you are unlocked. Most 8086Ks should be able to do 5 Ghz on more than one core.

 

[VirtualLarry]

There is a distinct trend in his posting and imho I question if even got either CPU and instead parroting things about the cpu he thinks he can negatively point out. Using an AMD cooler that you can't buy retail on a CPU that doesn't ship with a cooler is really straw that breaks the camels back there. You don't ebay one when there are better coolers for less (and never really mentioned how he got his).

I'm sure that Intel provided him the cooler, in his "subversive stealth marketing review kit".



/s (I think?)

 

[killster1]

300mhz!! wow this would be a big deal if it was a Pentium 2 or something. does his benchmarks differ from reviews? great trolling so funny bravo! i think the boosting on amd is crap too, wait for a load then boost if you ask me!