Intel's 14nm Process Smaller Pitch, Smaller Die, Same Power Consumption

[know of fence]

Intel never claimed that Broadwell or Skylake would bring huge improvements. Still, looking at the Reviews of i7-5775C, I was surprised to see the pundits almost unanimously recommend Devil's Canyon instead, because it was much more affordable, available and higher clocked. I for one was impressed of the 5775C eDRAM gaming prowess and low idle power.

After Skylake was released somewhere between the pages and pages of technobabble, the IPC talk, meaningless benchmarks, and neatly arranged oranges and apples one simple fact was lost: Power consumption and power efficiency didn't change at all.

Isn't at the end of the day Performance per Watt the ultimate CPU benchmark? Haven't we all, myself included always taken for granted that a new process will bring better efficiency? - Not the case with Intel's 14nm, it seems.

Only much later looking at curiously high OC voltages it became clear that (load) power consumption hasn't changed. The best/only point of comparison is the i7-4790K and the i7-6700K running at the same base clock when multi threaded.

​

The differences in DRAM, chipset, motherboard as well as from one CPU to the next can be significant. The Techreport also measured the energy to complete a particular movie as slightly smaller but effectively equal. The rest of these bar graphs don't really allow efficiency comparisons, since the CPU clock and voltage are the determining factors, and clocks vary between the CPUs.

Slightly lower energy to complete the task for Skylake may be indicative of improved IPC here, or it may be a consequence of Turbo behavior.

So Techreport made no effort to discuss Vcore or OC'ing. TweakTown(dot)com on the other hand have been doing clock-for-clock comparisons for a while by setting all CPUs to 4GHz. Let's just assume IPC is close to equal. TT's 14nm Broadwell review shows no efficiency improvement, their measuring of power only works reliably for CPUs with integrated Voltage Regulators (FIVR).
"Broadwell and Haswell both have an integrated voltage regulator and all CPU power comes from the 8-pin. However, Ivy Bridge has some of its power coming from the 24-pin as well for the other voltage rails other than CPU and iGP. "

​

This Graph shows at least what kind of variation we can expect, from the same process, yet differently binned CPUs. Almost 18 W between early Haswell and 4790K both running at 4GHz. Also shows Broadwell behind the 4790K. The opposite of increased efficiency from a die shrink.

From their Skylake testing...

Curiously TweakTown's reviews doesn't explain if and how the VRs on Skylake are supplied. But power is still only measured at the 8pin CPU connector. Does the 6700K get also power from the huge motherboard connector?
I don't know if i should trust their Skylake number for these reasons, it might be that they just got a crazy good Sample, stating "I only needed 1.35v to get 4.8GHz under AIDA64 for an hour, and my temperatures didn't go above 80C [...] My really good Sandy Bridge CPUs could overclock and grab a maximum validation of 5.2GHz at 1.5v, Skylake can do the same. For reference, my good Ivy Bridge and Haswell CPUs had a very hard time doing this." - Source

So anyway TT has nothing but good things to say about the Intel CPUs, even though their Broadwell wattage shows no efficiency improvement and their Skylake thing is clearly either a statistical aberration or doesn't account for all of the motherboard power. Anandtech regulars should be no strangers to heaving to read between the lines.

As of recently best power consumption testing is probably done by Tom's and their 4790K (78.9 W) to 6700K (100.4 W) torture testing has 14nm consuming 21 W more. However as often the case with students and testers the author doesn't trust his data and he is careful not to draw conclusions. It hasn't occurred to him, that the old assumptions about improved power every shrink don't apply. So he promises to retest those numbers if required. Their CPU wasn't that great either: reaching an unstable 4.9 GHz @ 1.41 V. - Source

Intel have said that leakage was reduced and that the 14nm process can actually handle higher voltages, something quite unprecedented. IMO it is a very important fact and a necessary perspective for everyone. I'm not saying that jet fuel can't melt steel beams, but maybe we all should take a look at Skylake voltages and power. Intel were careful to say that we wouldn't see a jump like with the introduction of 22nm and that 14nm is a refinement of their previous process. Will the process mature over time, will the Kaby-lake refresh just bring higher clocks?
Power consumption differences are more pronounced in 4 GHz desktops, but they are proportional and they have a bigger impact on battery powered devices. MY (edit: wrong) guess would be that there too, 14nm hasn't claimed higher base clocks at the the TDP targets of 15 W and 45 W. Although clearly there have been improvements to idle power and power management (edit: below 3.3 GHz there seem to be power consumption gains troughout). I'll have to look into that next.

 

[ShintaiDK]

You can just read the package power of the CPU.

This is the power usage of my 6700K running CB11.5 at 4Ghz.

 

[The Stilt]

Unlike with Haswell and Broadwell, Skylake has no FIVR so you can quite easily measure the actual power consumption. A high-end motherboard VRM reaches ~88% peak efficiency so if you use 85% as the ball-park value, you´ll get pretty accurate estimation.

The power figures measured for Haswell and Broadwell are inflated, since there is a second voltage conversion after the main VRM (VRIn). IIRC Intel has never given any official numbers for the FIVR efficiency, however the early demos illustrated test data with 78% efficiency. The efficiency is expected to be quite poor since FIVR must operate at ultra high frequencies (inductance and capacitance restrictions, due physical size). FIVR operates around 466 times faster than the conventional VRM located on motherboard (140MHz vs. 300kHz) on average.

If you multiply the power consumption of Haswell & Broadwell by ~ 0.8, you´ll get a figure which can be better compared to parts which are not equipped with FIVR.

 

[jpiniero]

I wouldn't be surprised if the voltage needed at 4+ Ghz hasn't changed much.

You can just read the package power of the CPU.

This is the power usage of my 6700K running CB11.5 at 4Ghz.

Measuring at the wall is still the best way to measure power consumption even if it makes getting the actual CPU power draw hard.

 

[NTMBK]

I believe 14nm's efficiency improvements come further down the frequency curve. Mobile parts get a nice boost, 4GHz desktops not so much.

 

[frozentundra123456]

I believe 14nm's efficiency improvements come further down the frequency curve. Mobile parts get a nice boost, 4GHz desktops not so much.

+1

 

[Yuriman]

+2

Look at the power consumption of the 6700 vs the 4770. Skylake seems to do worse above 4ghz, but better at lower frequencies.

 

[Abwx]

I believe 14nm's efficiency improvements come further down the frequency curve. Mobile parts get a nice boost, 4GHz desktops not so much.

Barely 15 % better perf/watt at regular frequencies..

I don't know if i should trust their Skylake number for these reasons, it might be that they just got a crazy good Sample, stating "I only needed 1.35v to get 4.8GHz under AIDA64 for an hour, and my temperatures didn't go above 80C [...] My really good Sandy Bridge CPUs could overclock and grab a maximum validation of 5.2GHz at 1.5v, Skylake can do the same. For reference, my good Ivy Bridge and Haswell CPUs had a very hard time doing this." - Source

So anyway TT has nothing but good things to say about the Intel CPUs, even though their Broadwell wattage shows no efficiency improvement and their Skylake thing is clearly either a statistical aberration or doesn't account for all of the motherboard power. Anandtech regulars should be no strangers to heaving to read between the lines.

That s not possible, 14nm need more voltage at the same frequency than 22nm, even at 2GHz, check the existing numbers...

Intel have said that leakage was reduced and that the 14nm process can actually handle higher voltages, something quite unprecedented.

Indeed since lower leakage is due to higher threshold voltage, actualy perf/watt improvement comes exclusively from reduced capacitance, the transistor itself has lower transconductance wich mandate higher supply voltage.

 

[know of fence]

Unlike with Haswell and Broadwell, Skylake has no FIVR so you can quite easily measure the actual power consumption. A high-end motherboard VRM reaches ~88% peak efficiency so if you use 85% as the ball-park value, you´ll get pretty accurate estimation.

The power figures measured for Haswell and Broadwell are inflated, since there is a second voltage conversion after the main VRM (VRIn). IIRC Intel has never given any official numbers for the FIVR efficiency, however the early demos illustrated test data with 78% efficiency. The efficiency is expected to be quite poor since FIVR must operate at ultra high frequencies (inductance and capacitance restrictions, due physical size). FIVR operates around 466 times faster than the conventional VRM located on motherboard (140MHz vs. 300kHz) on average.

If you multiply the power consumption of Haswell & Broadwell by ~ 0.8, you´ll get a figure which can be better compared to parts which are not equipped with FIVR.

The point is to compare the impact of the manufacturing process, and even though Broadwell isn't just a die shrink of Haswell, it doesn't seem to have improved by much / at all. It's also bested in efficiency by the 4790K. This also shows that cherry picked low voltage parts gain more efficiency (4770K -> 4790K) than a 14nm die shrink.
Hmmm, I guess it also means that without Devil's Canyon/Refresh we would see a kind of "traditional" improvement.

If the power numbers are inflated because of FIVR this would make Skylake look even worse in the 4790K to 6700K at-the-wall consumption comparison. It is a third factor to explain the TweakTown 79 Watt figure, along with lucky* binning and a potential further power source. Do you know whether the 8pin CPU connector is the only source of CPU power for Skylake?
I don't really know how efficiency conversion works going from 12V to Vccin to Vcore, but they only bumped TDP by 9% going from Ivy to Haswell.
Also the explanation was, we moved VR to the CPU, not we added another VR with limited efficiency.

 

[Arachnotronic]

Barely 15 % better perf/watt at regular frequencies..




That s not possible, 14nm need more voltage at the same frequency than 22nm, even at 2GHz, check the existing numbers...




Indeed since lower leakage is due to higher threshold voltage, actualy perf/watt improvement comes exclusively from reduced capacitance, the transistor itself has lower transconductance wich mandate higher supply voltage.

Here we go again...

 

[ShintaiDK]

Here we go again...

 

[The Stilt]

The point is to compare the impact of the manufacturing process, and even though Broadwell isn't just a die shrink of Haswell, it doesn't seem to have improved by much / at all. It's also bested in efficiency by the 4790K. This also shows that cherry picked low voltage parts gain more efficiency (4770K -> 4790K) than a 14nm die shrink.
Hmmm, I guess it also means that without Devil's Canyon/Refresh we would see a kind of "traditional" improvement.

If the power numbers are inflated because of FIVR this would make Skylake look even worse in the 4790K to 6700K at-the-wall consumption comparison. It is a third factor to explain the TweakTown 79 Watt figure, along with lucky* binning and a potential further power source. Do you know whether the 8pin CPU connector is the only source of CPU power for Skylake?
I don't really know how efficiency conversion works going from 12V to Vccin to Vcore, but they only bumped TDP by 9% going from Ivy to Haswell.
Also the explanation was, we moved VR to the CPU, not we added another VR with limited efficiency.

All of the CPU power, regardless of the VRM configuration come from EPS12V and / or ATX12V connectors. Very few consumer motherboards have both of them, so unless the measurements are botched badly the numbers measured using these connectors are accurate.

FIVR requires 1.5 - 1.8V high current input and that voltage is generated by the motherboard VRM from the 12V input. Since FIVR cannot use 12V input directly there are additional conversion steps required, which decrease the total efficiency. Until it can use 12V input directly it doesn´t really simplify the motherboard design much and will have a negative effect on efficiency.

 

[know of fence]

I believe 14nm's efficiency improvements come further down the frequency curve. Mobile parts get a nice boost, 4GHz desktops not so much.

Looking at the 15W TDP cpus, this seems to be the case indeed. 14 nm core-i7 SKUs get a massive 0.5 GHz base frequency boost.

http://ark.intel.com/compare/76616,85215,88192

So I guess we have a frequency curve situation like this. (taken from an unrelated PCper Nvidia review)

You can just read the package power of the CPU.
This is the power usage of my 6700K running CB11.5 at 4Ghz. Package power ->57.64 W

If you give Vcore as well, we could have a first point of comparison for package power values.

 

[coercitiv]

So I guess we have a frequency curve situation like this.

Indeed, that should be the case.

However, there's one more thing to point out: we have quite a lot of variation in voltage supplied to CPUs from mobo manufacturers. Looking at the Skylake review on Anandtech I was quite surprised by the delta between stock voltage and "manual undervolt". A voltage margin of 150-250mV or even more (stock 4200Mhz vs. undervolted 4300Mhz) is quite surprising to say the least.

I believe ShintaiDK also mentioned a drop in stock voltage after a recent BIOS update on his rig.

Granted, I've been out of the desktop game for quite a while and I've grown used to the minuscule voltage margins achieved on mobile Haswell (40-50mV adaptive offset is usual of HW-U, with values increasing toward 100mV for the rest of mobile parts). Maybe someone who knows desktop Haswell better can shed some light here, I may be completely off base.

To be clear, I'm not saying the voltage margin eats the entire power advantage, just that it paints a darker picture.

 

[ShintaiDK]

I believe ShintaiDK also mentioned a drop in stock voltage after a recent BIOS update on his rig.

Correct, 1.240V dropped down to 1.168V give or take.

 

[know of fence]

Indeed, that should be the case.

However, there's one more thing to point out: we have quite a lot of variation in voltage supplied to CPUs from mobo manufacturers. Looking at the Skylake review on Anandtech I was quite surprised by the delta between stock voltage and "manual undervolt". A voltage margin of 150-250mV or even more (stock 4200Mhz vs. undervolted 4300Mhz) is quite surprising to say the least.

I believe ShintaiDK also mentioned a drop in stock voltage after a recent BIOS update on his rig.

Granted, I've been out of the desktop game for quite a while and I've grown used to the minuscule voltage margins achieved on mobile Haswell (40-50mV adaptive offset is usual of HW-U, with values increasing toward 100mV for the rest of mobile parts). Maybe someone who knows desktop Haswell better can shed some light here, I may be completely off base.

To be clear, I'm not saying the voltage margin eats the entire power advantage, just that it paints a darker picture.

For some time I've been running a -169 mV adaptive offset undervolt on a completely mediocre, or perhaps below average OC'ed Haswell Pentium AE.
It makes sense to under-volt on the desktop for users (not really for manufacturers), even more so for laptops. So it's not surprising that Auto voltage is tuned closer to minimal values on mobile parts, where battery life is often tested and advertised.

 

[sm625]

Do not forget that temperature has an impact on power consumption. Could it be that 4790K's superior IHS attachment leads to lower temperatures which in turn leads to lower power consumption? What you really need to do is delid all of these chips if you really want accurate power efficiency measurements of the process.

 

[know of fence]

I have the load wattage numbers from hardware.fr, Techspot and Guru3D lined up leaning either one way or the other. My guess is if I compiled a large enough list and took the mean, the difference would be around 3W in favor of Devil's Canyon (88W), which is the difference in both the Processors' TDP defined at their base frequency of 4.0 GHz.

It would be great to see actual voltage frequency curves of course, I made one for Haswell. Regardless of if you have a good SL, BW or a bad one, even comparing two randomly picked CPUs should produce different slopes for the linear part of the curve and confirm that 14nm is really more of a low power process. Something that high TDP and low Turbo are indicating. But setting a higher TDP and allowing higher voltages could just as easy be a way for Intel to improve yields.

 

[know of fence]

This topic is almost exhausted, though not resolved until actual proof in the numbers and the pudding.

This Broadwell slide touts Performance per Watt improvement for 14nm, but not for Desktop. (just Server, Laptop, Mobile) anything < 3.3 GHz

This slide actually states that 14nm "was optimized for low-voltage performance" for the mobile BDW-Y process variant. It may not be for Skylake?

​

Do not forget that temperature has an impact on power consumption. Could it be that 4790K's superior IHS attachment leads to lower temperatures which in turn leads to lower power consumption? What you really need to do is delid all of these chips if you really want accurate power efficiency measurements of the process.

How many additional Watts will a quad-core consume a running stress test 5 to 20 °C hotter. I don't know to be honest.

[edit] Never forget IDC's epic topic: Sandy Bridge at very high voltage still only needs 15 W more (at the wall) for ~20°C delta.

 

[el etro]

Giving credit to Intel's 14nm: It turned 15W chips decent for the first time. Proofs it that 28W-35W Dual-Core processors practically disappeared from Broadwell and on.

 

[IntelUser2000]

Looking at the 15W TDP cpus, this seems to be the case indeed. 14 nm core-i7 SKUs get a massive 0.5 GHz base frequency boost.

These are the reasons that cement my belief that we are at a very end of process scaling and CPU advancement in a traditional way.

Those 15W CPUs, while the base clocks are higher, for most of users only the Turbo clocks matter, as reflected by benchmarking. The sustained Turbo is better too, but the peak Turbo has stayed same, or even dropped. Could be an indication there's a problem with reaching high frequencies.

confirm that 14nm is really more of a low power process. Something that high TDP and low Turbo are indicating. But setting a higher TDP and allowing higher voltages could just as easy be a way for Intel to improve yields.

For high frequency CPUs: Yield is a very good reason IMO. If you want to make certain numbers of high frequency chips, and you can't, increasing voltages may allow you to reach yield targets. Likely a reason why the characteristic "bad" chips like Prescott, NV30, Bulldozer all ran especially hot compared to how "fast" they ran.

For low power CPUs: What about Atom? Airmont turned out mediocre. We were supposed to have 2.7GHz chips! This, who knows what happened.

 

[NTMBK]

For low power CPUs: What about Atom? Airmont turned out mediocre. We were supposed to have 2.7GHz chips! This, who knows what happened.

Atom uses a different process variant, the SoC process. That may well be what borked the performance, it is optimized for density instead of performance.

 

[know of fence]

Fellow forum member Dufus kindly provided me with the voltage Data about his Notebook BW core i7-5500U after my request. According to him the VIDs were read from CPU registers using some kind of programming magic. His values are compared to the Speedstep Auto voltages of my G3258 which were read from CPU-Z forcing a power state with windows power settings. I used only data below base frequency to keep things straight and simple.

The position of the Voltages on the Y-Axis isn't really saying much, because here we compare a cheapo Pentium with a ULV-mobile-CPU. What matters is the slope or angle for the Voltage - Frequency - Curve.
IMHO this confirms and demonstrates that Broadwell uses a 14nm low-power-process, at least for this 15 W CPU.
The BW-straight has a slope that is 26% steeper than that of HW, meaning that ascending multipliers require a voltage bump that is by x1.26 higher than those of HW.
I chose red and green based on this graph posted earlier.

 

[Ferzerp]

You've posted some nice power draw graphs, but they don't do anything to touch on performance.

You can't discuss performance per watt with only half the data.


edit:

Going to leave this, but after I typed it, I saw the kJ graph.

 

[crashtech]

Looking at the OP, it seems that if the data were normalized for clock speed, the results might look a bit different.