Why does Intel care less about power-saving on the desktop than mobile?

[Revolution 11]

Intel probably cared too much about power consumption for Haswell. That's probably the main reason why clock speeds for Haswell have remained so low.

That and other reasons like vastly increased power and heat requirements as you pass from 4.0-4.2 Ghz and move past 4.5 Ghz. There is a reason that Intel clocks have hovered around 3.4-3.6 even for their best processors, that is the most efficient clock-rate for Intel's process.

Now obviously you can try for higher clocks as the default spec for a process node. Both Intel (Pentium 4) and AMD (Bulldozer) have tried for a high-clock design but the results were not good.

 

[tential]

Then you could just buy a mobile processor SOC and use it. That's basically what you'd get anyway. Look at the J1900 which is kind of a hybrid, mobile tech but in a desktop form factor.

Yup this. If you want an efficient chip it's not like Intel isn't sellling them. And it's not like they aren't selling them cheap.

I'd say a TON of people could get by on a J1900. Most of the tasks I'm doing right now I could do on a J1900.

 

[raildogg]

On the other hand, why do we care so much about this? Is it the minuscule energy savings? Or something more?

 

[Phynaz]

The rig in question is my HTPC, and it's a G1610 in an ECS H61 mini-ITX mobo, with a 120W power-brick (like a laptop one, but a bit bigger).

Those power bricks are horribly inefficient.

 

[Homeles]

On the other hand, why do we care so much about this? Is it the minuscule energy savings? Or something more?

As evident pointed out, when you ship millions of units, even saving a small amount adds up.

 

[evilr00t]

The Ivy chips do run at 800MHz, but you'll have to use software like ThrottleStop to get there. I think this is the MAX_EFFICIENCY_RATIO setting in the CPU MSRs, which is set to x16 on the Desktop SKUs and x12 on the mobile SKUs (at least, this is what my 3740QM does). This doesn't change the minimum multipler.

In practice, I don't see a power consumption difference between x8 and x16 on desktop when idle. The guess here is that all the time spent idle at 800 and 1600 are in C-states, so yeah. If you're loading up the cores at 800 vs. 1600, yeah you will see a difference.

Just another piece of interesting information is that the Quad Core Ivy Bridge CPU I have in my laptop reaches the most efficient clock speed around 1.8-2.1GHz. This is cores*clock/ia32_power, which climbs up rapidly from 800 to 1600, levels off around 1900, and drops off slowly after 2200.

 

[ninaholic37]

Just another piece of interesting information is that the Quad Core Ivy Bridge CPU I have in my laptop reaches the most efficient clock speed around 1.8-2.1GHz. This is cores*clock/ia32_power, which climbs up rapidly from 800 to 1600, levels off around 1900, and drops off slowly after 2200.

Interesting. So for the best battery life and lowest amount of heat, it should idle at 800mhz and raise up to 1900mhz under full load? How can I test this on my Pentium M 750?

 

[unclewebb]

Ivy desktop CPUs can use the 8 multiplier and get down to 800 MHz but the question is, why would anyone want to do this?

With Intel CPUs, for maximum power savings, all you have to do is enable the low power C States. When idle, individual cores should be spending 99% of their time in one of the deep C States like C6 or C7 for Haswell CPUs. When a CPU has idled down like this, the MHz becomes a meaningless number.

In the above pic my entire computer was consuming 49 Watts at the wall as measured with a Kill-a-Watt meter. The same computer when cranked up to 4400 MHz consumes the exact same 49 Watts when idle. Why? In both situations the CPU cores are spending 99% of their idle time in C6.

Intel likely blocked the 8 multiplier in these CPUs because it is pointless. There is nothing to be gained by running a modern Intel CPU this slow. The advantage of a fast CPU when idle is the CPU can process all of the various Windows background tasks quickly. This allows the CPU cores to spend a larger percentage of time in one of the deep C States where power consumption is minimized.

Some old school methods of saving power are obsolete. Before trying to make your CPU slow and inefficient, try doing some power consumption testing.

Edit: Here's a Core i7-4700MQ Haswell CPU with the cores spending 99% of their idle time in C7. MHz at idle is unimportant compared to how much time your CPU cores and CPU package are spending in the low power C States.

 

[Lepton87]

My computer idles at 110-130W depending on the power policy and I don't care, it's the same amount my monitor draws or my room lighting draws.(probably more, I have LED bulbs and conventional bulbs I hate those energy saving bulbs. I don't like the light they emit. They are the most energy efficient but I don't use them in my room anyway. I use them elsewhere in the house though.

What program do you use for that?

 

[TuxDave]

Edit: Here's a Core i7-4700MQ Haswell CPU with the cores spending 99% of their idle time in C7. MHz at idle is unimportant compared to how much time your CPU cores and CPU package are spending in the low power C States.

Thanks for saying what I wanted to say. You have a very narrow window in "idle" where the clocks are running but the core hasn't shut down.

 

[unclewebb]

What program do you use for that?

RealTemp T|I Edition
https://www.sendspace.com/file/55yvry

The C State window above is from RealTemp 4.00 which has not yet been publicly released. The RealTemp T|I edition also includes core and package C State reporting. Intel's Core i CPUs include high performance timers so it is easy for software to accurately measure C State data.

 

[jpiniero]

That's the thing - 49 W for idle is way too much. I'm sure there's more Intel could do to improve on that, but surely the rest of the system would have to be improved to get that number down.

 

[Lepton87]

RealTemp T|I Edition
https://www.sendspace.com/file/55yvry

The C State window above is from RealTemp 4.00 which has not yet been publicly released. The RealTemp T|I edition also includes core and package C State reporting. Intel's Core i CPUs include high performance timers so it is easy for software to accurately measure C State data.

Thanks, I've been using RealTemp but 3.6 version. I noticed a new reading, package temp.

 

[CHADBOGA]

That's the thing - 49 W for idle is way too much. I'm sure there's more Intel could do to improve on that, but surely the rest of the system would have to be improved to get that number down.

I'm sure we will be seeing Intel improve greatly in that area anyway in the next few years.

 

[Torn Mind]

That's the thing - 49 W for idle is way too much. I'm sure there's more Intel could do to improve on that, but surely the rest of the system would have to be improved to get that number down.

The 49 watts are most likely from the mechanical hard drives and other peripherals. My i7-3770S is pulling about 10 watts at idle according to HWmonitor. I don't have a watt meter to measure the pull of my system at the wall, but the Haswell CPU's power consumption should be lower than my 3770S.

I think when Anand first review Haswell, idle consumption at the wall was 10 watts lower than the previous Ivy Bridge idle results. He did mention that the result was more than he expected from Intel alone, though.
http://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/2

 

[Homeles]

Intel probably cared too much about power consumption for Haswell. That's probably the main reason why clock speeds for Haswell have remained so low.

There's a lot of truth to that, although the matter of it being "too much" is subjective, of course. If desktop CPUs are the only thing one cares about... yeah, it's can be tough to swallow.

The IVRs are undoubtedly the primary reason behind Haswell's higher TDPs compared to many of the IVB SKUs. IVB has a TDP of 77W, while Haswell has a TDP of 84W. Real world power consumption, even at the system level where a lot of the system board level VR circuitry is reduced, is generally is higher under heavy load as well.

Even assuming a 90% efficiency, this already puts Ivy Bridge at Haswell's power level (77W / 0.90 = 85.5). TDPs are coarsely related to power consumption of course, but the effect can be seen clearly. The Hillsboro team must have done a lot of work to keep power from being higher than it ended up being, despite the integration of the VR, especially when you consider the expansion of the GPU.

Then you factor in that the IVRs are nestled up against the "long" sides of each core:

...and there goes your overclockability.

The implications for mobile however are very good. Board size goes down significantly (PDF download warning), as does the bill of materials by about $5 (which can be applied to desktop boards as well). And as I'm sure you know, battery life can be as much as 50% higher with Haswell. The U-series and Y-series SKUs benefit most from this.

There are benefits that the IVR brings to overclocking, like the extremely clean power being supplied, but they're significantly outweighed by the fact that there's more power and heat being released, and at close proximity to the CPU cores. Unless these can be built with near 100% efficiency... these IVRs are probably always going to hold back max overclocks. Still, there is a long way to go in the realm of integrating power supply components on die.

However, in my opinion, it was a change for the better, overall. I don't know the whole story behind Skylake's exclusion of the IVR, but from everything I've seen from Haswell, it wasn't the right decision. It'll be awesome when Atom gets inevitably gets FIVR, just like it's awesome for today's Haswell-powered mobile devices.

That's the thing - 49 W for idle is way too much. I'm sure there's more Intel could do to improve on that, but surely the rest of the system would have to be improved to get that number down.

A lot of that is dependent on the motherboard being used. I highly doubt that the majority of that 49W is from the CPU. In fact, it's the dGPU at fault, if he's got one installed.

Normally Haswell systems will idle in the 20W range with a decent board. The processor can actually get to 45mw in C10. Even AMD's previous generation Richland processor idled at ~1.6W. Haswell should be well south of that. In fact, C6/C7 calls for a minimum 0.05A of power on the 12V2 rail, which is 600 mW. Even Ivy Bridge had a C7-state TDP as low as 2.2W. I think it's pretty clear that it's not Intel's hardware that's responsible for the 49W idle.

 

[unclewebb]

That's the thing - 49 W for idle is way too much.

That includes an Nvidia GPU and my Core i5 is overclocked to 4400 MHz. 49 Watts is less than half what my similar Core 2 Quad QX9650 system consumes when idle so Intel and computer manufacturers are definitely heading in the right direction.

Here's some math.

50 Watts = 0.05 kWh

If my computer is idle for 8 hours a day for 30 days a month.

8 hours x 30 days x 0.05 kWh = 12 kWh per month

Where I live, power costs about 11 cents per kWh.

Monthly cost = 12 kWh x $0.11 = $1.32

If computer manufacturers can cut power consumption in half, I will save 66 cents a month. That is probably the reason Intel does not enable the low power C8, C9 or C10 C States on their desktop CPUs. For individual desktop users, the savings are minimal.

Lepton87 - Intel added an extra register to their newer CPUs that contains CPU package temperature information. The individual core temperature sensors as well as the Intel GPU temperature sensor and a few more on die temperature sensors are all compared internally and then the highest temperature is reported as the CPU package temperature. With multi-core CPUs, this makes it a lot easier. You only need to keep an eye on a single temperature number.

 

[ninaholic37]

Keeping my CPU at 1867mhz (highest speedstep) always makes my laptop much hotter than at 800mhz (lowest speedstep). I think I only have up to C4 though, and not even sure that C2-C4 work right, or do anything useful when my (single core) laptop is on. Old school "obsolete" methods FTW!

 

[Homeles]

If computer manufacturers can cut power consumption in half, I will save 66 cents a month. That is probably the reason Intel does not enable the low power C8, C9 or C10 C States on their desktop CPUs. For individual desktop users, the savings are minimal.

It's because those power states wouldn't really ever be used on a desktop. They're for connected standby, which is basically what a phone or tablet would go into when the you turn off the screen. The device is still on, but it's not in use. It's not powering the display. It's not crunching numbers. It's just waiting for some sort of external stimulus, like an incoming email.

I think you can see why this is nearly pointless for a desktop. If you're not using it, the entire computer will essentially go to sleep (S3, or suspend to RAM), which is a deeper idle state than C states (still S0). If your screen is powered on, but your PC isn't doing anything in particular, it's going to be going down to C6 or C7.

As far as power savings go, you'd be fairly close to half of your machine's idle power if you were on a Haswell platform (Anand got about 33% reduction in his HSW review). There are a lot of reasons for this aside from the obvious architectural improvement. For example, the PCH is moved to 32nm from 65nm. Motherboards have also gotten a lot better when it comes to idle power, in general.

 

[Torn Mind]

Keeping my CPU at 1867mhz (highest speedstep) always makes my laptop much hotter than at 800mhz (lowest speedstep). I think I only have up to C4 though, and not even sure that C2-C4 work right, or do anything useful when my (single core) laptop is on. Old school "obsolete" methods FTW!

I used a Dell D810 laptop with that chip myself. You have to keep it at 800 MHz because at 1.86 GHz, the battery life is halved at the very least and the CPU can trigger the fan to fire up.

Those Pentium Ms were one of the precursors to the high performance chips that didn't guzzle down a lot of power we have nowadays. Back then, the Pentium Ms at low clockspeeds matched Pentium 4s at really high clockspeeds.

 

[jj109]

There are benefits that the IVR brings to overclocking, like the extremely clean power being supplied, but they're significantly outweighed by the fact that there's more power and heat being released, and at close proximity to the CPU cores. Unless these can be built with near 100% efficiency... these IVRs are probably always going to hold back max overclocks. Still, there is a long way to go in the realm of integrating power supply components on die.

IVR and overclocking is a big question, imo. Just how clean is that power when CPU loads approach 150W, 200W, 250W (AVX2)?

 

[Homeles]

IVR and overclocking is a big question, imo. Just how clean is that power when CPU loads approach 150W, 200W, 250W (AVX2)?

Probably cleaner than it'd be if that power were supplied off-package.

 

[ZGR]

The i7-3720QM in my laptop idles at 1.2 GHz unless I use Throttlestop. The power savings are not really noticeable and neither are temperatures going down to 800 MHz. I think this is because the voltage does not get below 0.7855 even while idling all the way to 1.6 GHz.

I would be interested to see power consumption saved over time as 800-1600 Mhz is perfectly useable for basic tasks, keeps the CPU at minimal voltages, and can be passively cooled. If there is a large battery life impact difference from 800 MHz to 1600 MHz for a light web browsing test, then I would be interested. For the most part, getting tasks done faster is far more important to me. But in situations where I need all the battery life I can get, Turbo Boost can be very detrimental.

 

[Lepton87]

There's a threshold voltage below which a CPU won't work and I don't think that it's different between 800MHz and 1600MHz, at the same voltage the CPU clocked at 800MHz would consume half the power than that clocked at 1600MHz but it would take twice as long to complete the same task resulting in 0 gain in efficiency thus only making your computer slower.
ps. I'm talking about CPUs from the core line not mobile ARM CPUs, which obviously can operate at lower voltage at 800MHz then 1600MHz.

 

[BSim500]

There's a threshold voltage below which a CPU won't work and I don't think that it's different between 800MHz and 1600MHz, at the same voltage the CPU clocked at 800MHz would consume half the power than that clocked at 1600MHz but it would take twice as long to complete the same task resulting in 0 gain in efficiency thus only making your computer slower.

This is pretty much true. I just tested 800MHz vs 1600MHz on an Ivy Bridge via ThrottleStop - voltage doesn't go any lower automatically, and can't really be forced much lower below 0.65v via offset undervolt without experiencing instability at higher Speedstep freqs. With a kill-a-watt plugged in there was literally zero difference in 800MHz vs 1.6GHz idle power consumption both at any one time and over time, but the forced 800MHz often idle'd at 1000-1100MHz with a web browser running in the background whereas the 1.6GHz stayed at 1.6GHz. So overall, it's already as low as it can go. Likewise, the "UnCore" part of the CPU (L3 cache, memory controller, etc) will draw a fixed amount of juice anyway independently of SpeedStep, so halving freq won't exactly halve power consumption (core gets pulled off of VCore, uncore feeds off of VTT). Same is true of running the iGPU (another fixed minimum load within the CPU), which is partly why dual-cores are 55w TDP not 42w (exactly half of quad's 84w)).

Edit: Likewise for load power, a maxed out 4-core x264 encode draws only 33% more juice vs 2-core (on my rig anyway), yet provides over 80% higher fps.