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

[Yuriman]

I have a thread that analyzed Ivy Bridge (also 22nm):

http://forums.anandtech.com/showthread.php?t=2443083&highlight=

I'd be happy to participate. I'm thinking something along the lines of, finding the minimum voltage at a given clock (to rule motherboard overvolting out) and then measuring the delta idle/load at 200mhz intervals, probably from 1600mhz up to ~4.6ghz. Power supplies will make a difference here (and introduce some systematic error), but we can minimize it if anyone participating has a platinum-rated PSU.

 

[know of fence]

I have a thread that analyzed Ivy Bridge (also 22nm):

http://forums.anandtech.com/showthread.php?t=2443083&highlight=

I'd be happy to participate. I'm thinking something along the lines of, finding the minimum voltage at a given clock (to rule motherboard overvolting out) and then measuring the delta idle/load at 200mhz intervals, probably from 1600mhz up to ~4.6ghz. Power supplies will make a difference here (and introduce some systematic error), but we can minimize it if anyone participating has a platinum-rated PSU.

I'll have to come up with a sensible template and need to do some testing first.

I'm afraid probing the voltage curve for your CPU may take a lot of time and stability testing. And the numbers that you will produce will be called into question: Is it really stable with Linpack, OCCT, prime95 and for what duration.

But I'm convinced that it's at least worth it to determine the minimum stable Vcore for CPU's base frequency and your OC of course.
The difference between Auto and manual min Vcore at base clock determines just how low an offset undervolt can be set. (I've set a - 183 mV Undervolt this way in my old post, later corrected to -169 mV)

I would look at Auto Voltages for starters, because it's easily done entirely from Windows, by adjusting the power state. The Spedstep voltage curve for CPUs with Turbo isn't a straight either, which is interesting in itself.
I'll have to test if package power displayed in HWinfo64 correlates to the at-the-wall Watt-meter reliably. So maybe package power is the easy first thing to compare.

However there is little to be learned from idle power or the delta to idle power, other than how much energy the CPU happens to waste when idle. Imagine an Ideal CPU or ideal PC, it would have the idle power 0.

1-10 min default OCCT
Power State%|Multi|Vcore|Package Power|Kill-a-Watt|hottest core max Temp (over x minutes)

 

[Shaun_Brannen]

A good CPU design is very balanced and synergistic. Much of the individual components simply don't make sense without the design balance of the other individual components. For example, as you increase execution width you need more reordering resources, prefetching, memory disambiguation, etc. If you have too much of those things your net efficiency goes down instead.

It all has to come together in unison. And the overarching ideal for that unison will be dictated by power and area targets.

I don't think there's really an artificial gap. It's easy to compare Atom to Twister and say that it sucks. But they have different design targets because Apple is only putting the CPU in expensive, premium flagship products. So they can afford to have something bigger, they can afford to limit themselves to two cores with big caches. These are things that those targeting the broad Android and Windows markets can't do, especially when, like Intel, they're also putting the chip in low end 28nm TSMC dies.

And it's not at all a given that Intel can do everything Apple does exactly as good or better. They have their own legacy of experience and their own bag of tricks. Intel are industry leaders in CPU design but their long term expertise has been in meeting laptop, desktop, and server targets while their lower power stuff has largely been taken for granted - in the UMPC and netbook days OEMs took what Intel would give them and like it because they were the only choice (or VIA, but they were never going to be real competitors) Only fairly recently has Intel really started building experience in aggressive low power CPU optimization and SoC integration. It was not something they could just become the best at overnight, even given their resources.

Silvermont was decent for when it came out and the 22nm node it was on. They could have and really should have been more aggressive in releasing a much better successor for 14nm. Unfortunately they're stuck in a tick-tock mindset with CPUs, where on the desktop side this looked extremely aggressive, but on the mobile side has been falling behind Apple and ARM. They need more continuous improvement to compete.

But I think they've lost motivation because even when they have a reasonably competitive SoC it's been hard to get really big wins with Android, especially phones. A lack of the connections other SoC makers have and being tied to x86 has held them back and now they might just feel like it's not worth trying anymore.

Agree 100%. I think Intel's just going through a bit of "fake it until you make it" right now. And I definitely don't think they're holding back artificially -- they're not that dumb. They know damn well that the mobile market is something they desperately need to get into.

 

[cytg111]

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.

just a moment to +1 this post "Informative".

 

[cytg111]

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)

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

I cannot find the post right now but i recall IDC posting about this as well, how the process (14nm) was engineered towards lower power, lower frequency too.

 

[coercitiv]

Some interesting info, at least for me: I installed Intel XTU on my Skylake system and it reported a +70mV CPU voltage offset while my BIOS settings were all on Auto voltage wise.

It was a rather odd finding, so I decided to do a bit of testing by enabling the Offset voltage mode through BIOS and setting the smallest offset possible. I set a +5mV offset, and the same +5mV value was reported via Intel XTU.

Yet the Vcore changed drastically:
Auto voltage @ 3.9Ghz -> 1.208V
+5mV Offset @ 3.9Ghz -> 1.16V

Turns out Auto settings on this MSI mobo enable a generous +70mV offset. I guess it's a safe way to make sure anyone can overclock their CPU by turning just one dial

 

[know of fence]

@coercitiv

Thanks for providing me with some numbers. I hoped the Auto voltage thing would be a nice clear cut experiment to start with, but it turns out that with the VRs back on the motherboard for Skylake, things would not be simple.

• For starters 1xx motherboards don't scale Vcore further down than 0.800 V, as has been the case in the past with Sandy and Ivy Bridge. (not pictured). It's possible to go down to 800 MHz but Vcore remains the same.

• The other thing is that the Auto Voltage curve, isn't straight unlike that of 22 nm Haswell. Rather both 14 nm CPU's curves are comprised of two differently sloped straights, and curiously a kink in the middle at 2.5 GHz. Maybe that is where the power curves intersect as well.

​

• The good news is that the Skylake Auto-Voltage curve is different and more flat than that of Broadwell, which may indicate (speculation) that indeed a different variant of the process was used.
• Skylake's increased power consumption may not be just the result of a process optimized for low clocks, or high binning, but it may also be the case that the Cores also got "bigger" and more complex. Thus power consumption increases faster than its predecessors. The cores get hotter, faster. Bigger cores bring 9% IPC improvment in encode, and ~12% improvement in FPS in CPU and memory bound scenarios, although most of it is probably attributable to DDR4.

 

[Dufus]

@know of fence Haswell also has 2 slopes, one for non-turbo bins and one for turbo bins. Probably easier to see on mobile chips as there are usually a lot more turbo bins than on DT chips. Note also VID is only up to maximum default turbo bin, for unlocked bins above that, interpolation voltage is required.

page 19 - http://www.sstc.co.jp/products/Intel/SF13_AIOS001_100.pdf

page 5 - http://download.intel.com/support/m...z87klt75k_overclocking_assistant_june2013.pdf

 

[know of fence]

@ Dufus
It's indeed a weird coincidence that the kink is at the same place for both of these CPUs, a coincidence has nothing to do with the process node as such. Not having tested Turbo-CPUs myself I wasn't quite sure about the Turbo incline, but that Intel slides confirms it. For the Broadwell i7-5500U 2.4 GHz it is indeed where Turbo begins, as you are well aware.

But other than "Turbo kink", BW and HW Speedstep curves are straights, unlike SL and they scale way down to 0.550 V for negative offset HW and to 0.675 V default BW, which is one possible reason why Broadwell pulls ahead in many idle power consumption tests, vs Skylake with a set >0.800 Vcore.

At this point I consider it fact that Broadwell (including the i7-5775C) is indeed produced using 14nm Intel's low power process.

At default settings it idles at record lows, because it's a low power process that can also scale down Voltage a lot thanks to FiVR.

However when idle @ 4.5 GHz with an Override OC, it actually consumes
more than all of the predecessors. An OC'ed Desktop Broadwell even out-consumes AMD-FX in Far Cry 3 (Source).

 

[frozentundra123456]

What you say is true, but it is kind of apples to oranges to compare BW overclocked pretty much to the max to stock FX. Even then, the difference is probably close to the margin of error, and "overclocked" FX, aka 9590 still uses more power. BW at 4.5 is also the fastest of all in a lot of tests, and sometimes gives close to double the framerate of the FX.