[RWT]Adaptive Clocking in AMD’s Steamroller

[coercitiv]

Going back to the original article, it's not clear to me why the benefits of adaptive clocking are significant only for high frequencies (2Ghz+).

The effectiveness of this technique can be measured by the amount of voltage droop that can be bypassed, so the two most important factors in this mechanism should be effective voltage droop and adaptive clocking latency (detection+correction). Following this line of thought, two questions come to mind:

1. Does voltage droop get reduced at low frequencies by such amounts as to become insignificant from a power saving perspective?

2. Is adaptive clocking latency directly dependent to core frequency?

 

[Enigmoid]

Compere Richland 35W TDP A8-5550M which is a Quad Core 2.1GHz base and 3.1GHz Turbo against Core i5 4200U.
Kaveri A10-7500 has 2.1GHz base and 3.3GHz Turbo.

http://www.notebookcheck.net/AMD-A-Series-A8-5550M-Notebook-Processor.89639.0.html

http://www.notebookcheck.net/Intel-Core-i5-4200U-Notebook-Processor.93563.0.html


From my own tests i can tell you that Kaveri 7700K (3.4Base/3.8Turbo) has the same performance in Cinebench R15 with A10-5800K (3.8Base/4.2Turbo). It can sustain 3.7GHz Turbo (4 threads) when only the CPU is loaded.

So i believe the mobile 19W TDP A10-7500 will have the same or a little better performance than the A8-5550M 35W TDP Richland.

Even that 35W SKU drops clocks to about 2.4 ghz during MT CPU load.

From notebookcheck, review of notebooks with that CPU on your link.

Asus equipped the F550DP with an AMD A8-5550M APU based on the Richland architecture. The CPU part of the APU is a quad-core processor with a base clock of 2.1 GHz. The clock rate can be increased to up to 3.1 GHz (AMD Turbo Core). AMD specifies a TDP of 35 Watt for this APU.

The CPU clocks at 2.8 GHz during the single-thread tests of the Cinebench benchmarks. Time and again the clock rate of one core increases to 3.1 GHz. During the multi-threading tests, the CPU usually clocks at 2.4 GHz. However, the clock rate of all four cores frequently falls to 2.1 GHz. As a result, multi-threading applications cannot use the full potential of the CPU.

And

We can see that the clock fluctuates around 3 GHz in single thread applications, which enables a Cinebench R11.5 score of 0.78 points. Our review unit can actually beat the alleged top version A10-4600M (0.74 points), but is far behind a comparable Intel CPU like the Core i3-3110M (around 1.0 point).

The situation is slightly better in multithread scenarios, where all four processor cores of the modular architecture can be stressed. Although the clock drops from 2.8 to 2.4 GHz after 20 seconds, the AMD chip gets closer to the Intel competitor.

Even on a 35W tdp heavy MT CPU load forces clocks around 2.4 Ghz. Clock for clock Kaveri appears more efficient that richland/trinity but a 19W cpu only has around 54% the TDP. I am going out on a limb and guessing that under heavy MT load this 19W CPU will clock around 2.0-2.2 ghz at most.

 

[rainy]

Heard it about richland, heard it about kaveri (supposed to be 30% faster IPC on some of AMD's slides).

Sorry, it's simply not true: AMD have never promised to deliver 30 percent higher (not faster) IPC with Kaveri - they have talked about 30 percent higher ops.

These design improvements have resulted in a 30 percent reduction in i-cache misses, an increase of 25 percent on max-width dispatcher per thread and a 20 percent reduction in mispredicted branches, which results in a 30 percent increase in overall ops delivered per cycle.

http://www.tomshardware.com/news/AMD-Steamroller-Piledriver-Kaveri-processors,17217.html

Honestly, I don't know why you're using A10-4600M as example: are you really expecting that AMD haven't made any progress since releasing Trinity two years ago?

I am going out on a limb and guessing that under heavy MT load this 19W CPU will clock around 2.0-2.2 ghz at most.

Which means effectively no turbo or almost none in such situations in your opinion?

Btw, I don't think that many users will run heavy MT programs on 19W TDP APU.

 

[NostaSeronx]

Kaveri has the same-ish performance as Richland but with lower voltage and clocks.

 

[Piroko]

Going back to the original article, it's not clear to me why the benefits of adaptive clocking are significant only for high frequencies (2Ghz+).

The effectiveness of this technique can be measured by the amount of voltage droop that can be bypassed, so the two most important factors in this mechanism should be effective voltage droop and adaptive clocking latency (detection+correction). Following this line of thought, two questions come to mind:

1. Does voltage droop get reduced at low frequencies by such amounts as to become insignificant from a power saving perspective?

2. Is adaptive clocking latency directly dependent to core frequency?

Think at it from this perspective:
To switch a transistor from off to on you need a certain amount of charge. At low frequencies you will be able to achieve this even with (slightly) lower than expected voltage, because you have enough time to charge it. At high frequencies you have a lot less time, thus you need to keep a larger safety margin to compensate possible droop. (I really hope this isn't oversimplified, feel free to correct me)

My first thought was that this is an elegant solution that works well with a resonant mesh and saves die space that would otherwise be needed for decoupling capacitors. Might also make AMDs implementation of a FIVR much simpler.

 

[Enigmoid]

Sorry, it's simply not true: AMD have never promised to deliver 30 percent higher (not faster) IPC with Kaveri - they have talked about 30 percent higher ops.

http://www.tomshardware.com/news/AMD-Steamroller-Piledriver-Kaveri-processors,17217.html

Honestly, I don't know why you're using A10-4600M as example: are you really expecting that AMD haven't made any progress since releasing Trinity two years ago?



Which means effectively no turbo or almost none in such situations in your opinion?

Btw, I don't think that many users will run heavy MT programs on 19W TDP APU.

Ah...You are right. Though 30% higher operations per cycle is very very similar to instructions per cycle. Needless to say it was overinflated marketing talk (and yes intel and nvidia do this to).

I never mentioned the a10-4600m. That is simply a quote from notebookcheck.

No they may not but a heavier game can certainty use that weak CPU and put a lot of stress on the GPU. 384 gcn shaders at 500 mhz is going to substantially cut into a 19W budget. A moderate CPU load + heavy GPU load could easily be more power hungry than a heavy CPU load.

 

[DrMrLordX]

To be exact, those values are for 65W TDP and turbo is 3.8 GHz - for 45W you have 3.1/3.3 GHz.
http://www.cpu-world.com/CPUs/Bulldozer/AMD-A10-Series A8-7600.html

The A8 7600 at 45W is 3.1Ghz base, 3.3Ghz turbo. In 65W mode its 3.3Ghz base and 3.8Ghz turbo.

Hmm, looks like I need to read my source more carefully next time. Base clock of 3.1 ghz, but base turbo clock of 3.3 in 45W mode, with average turbo clocks of 3.4 ghz.

Regardless, it's sitting right on that 3.4 ghz sweet spot.

 

[AtenRa]

I am going out on a limb and guessing that under heavy MT load this 19W CPU will clock around 2.0-2.2 ghz at most.

Richland A8-5550M Base clock = 2.1GHz , Turbo = 3.1GHz

Kaveri A10-7500M Base Clock = 2.1GHz , Turbo = 3.3GHz

Im quite sure it will be able to sustain the base 2.1GHz in MT loads all the time.
Also i bet it will have at least the same or higher ST and MT performance(depending on the app and the laptop thermals) than A8-5550M in the majority of benchmarks.
So at the end, it will be very close to Core i5 4200U in CPU performance.

 

[ShintaiDK]

Hmm, looks like I need to read my source more carefully next time. Base clock of 3.1 ghz, but base turbo clock of 3.3 in 45W mode, with average turbo clocks of 3.4 ghz.

Regardless, it's sitting right on that 3.4 ghz sweet spot.

You cant get 3.4Ghz on stock with a 45W model. 3.3Ghz isnt base turbo, its max turbo with stock setting.

 

[DrMrLordX]

Well somebody sure needs to let PCMag know! Google seems to want to serve them up first on any search for the A8-7600's clock speeds.

 

[Exophase]

1. Does voltage droop get reduced at low frequencies by such amounts as to become insignificant from a power saving perspective?

Yes, because the dI/dt should be lower at lower frequencies. And lower still with inherently simpler/lower power CPUs.

 

[Enigmoid]

Richland A8-5550M Base clock = 2.1GHz , Turbo = 3.1GHz

Kaveri A10-7500M Base Clock = 2.1GHz , Turbo = 3.3GHz

Im quite sure it will be able to sustain the base 2.1GHz in MT loads all the time.
Also i bet it will have at least the same or higher ST and MT performance(depending on the app and the laptop thermals) than A8-5550M in the majority of benchmarks.
So at the end, it will be very close to Core i5 4200U in CPU performance.

Except it has a 19W envelope to play around in instead of a 35W TDP. I expect AMD to have improved turbo so ST should be similar or higher. ST will be way lower than the i5 (the i5 matches the a10-7850 in ST) but MT will be fairly decent. I expect that with both the CPU and GPU loaded (gaming) clocks will drop quite a bit.

What I like about AMD on mobile is that their CPU's are surprisingly power frugal. Not terribly efficient but decent idle power consumption. Kaveri is probably a nice step up though it will not be able to compete with haswell ULT.

Adaptive clocking is nice to see in Steamroller but it would also be nice for AMD/OEM's to lower the voltage on mobile CPU's in the first place. There is often quite a bit of fat to be cut there.