[RWT]Adaptive Clocking in AMD’s Steamroller

[Homeles]

The improvements in Vmin from adaptive clocking are greatest when the Steamroller core is operating at high frequency (up to 9% lower Vmin at 4GHz). High frequency circuits are much more sensitive to power supply noise and stand to benefit more as a result. Figure 2 shows the benefits of the adaptive clocking system on power consumption for a desktop productization of Steamroller, with savings as high as 19%.

Figure 2. Power benefits from adaptive clocking

The cost of implementing adaptive clocking is quite low, especially given the significant power benefits. The area overhead for adaptive clocking is minimal; the droop detection and clock stretching circuits are approximately 0.2mm2, whereas a Streamroller module (i.e., two cores and a 2MB L2 cache) is 29.5mm2. The additional clock stretching circuitry does increase jitter by 0.5-1%, but only when the clock stretching is engaged and the core is already operating at a reduced frequency – so there is no impact on peak frequency.

Really interesting article.

http://www.realworldtech.com/steamroller-clocking/

 

[Idontcare]

Obvious downside being that the performance suffers because the processor is decreasing clockspeed while letting the voltage transient subside. Not a free lunch.

 

[Exophase]

Obvious downside being that the performance suffers because the processor is decreasing clockspeed while letting the voltage transient subside. Not a free lunch.

By no more than 0.2% in the worst case benchmark at 3.4GHz, yielding a 14% power consumption savings. When power limited that'd conversely allow an increase in nominal frequency, easily outweighing the amount lost. Seems like an obvious win.

Too bad this is much less relevant for lower power cores.

 

[jj109]

This is very neat. I wonder how it compares to Haswell FIVR.

 

[Idontcare]

By no more than 0.2% in the worst case benchmark at 3.4GHz, yielding a 14% power consumption savings. When power limited that'd conversely allow an increase in nominal frequency, easily outweighing the amount lost. Seems like an obvious win.

Too bad this is much less relevant for lower power cores.

Aren't they quoting a "7% stretch", meaning an increase in the clock cycle of 7%?

Seems like they'd save more power if they captured the impact of temperature on the shmoo plot and then dialed in Vnom based on operating temperature.

At least in that case they'd be saving power all the time by not continuously targeting the Vnom necessary to ensure operation at TJmax and Fnom.

 

[Exophase]

Aren't they quoting a "7% stretch", meaning an increase in the clock cycle of 7%?

The frequency skewing only kicks in when there's a current transient due to idle execution units turning on due to a change in instruction mix. The 0.2% figure is a measured worst case induced by one benchmark. The 7% would just be hit over short durations normally.

Seems like they'd save more power if they captured the impact of temperature on the shmoo plot and then dialed in Vnom based on operating temperature.

At least in that case they'd be saving power all the time by not continuously targeting the Vnom necessary to ensure operation at TJmax and Fnom.

That's orthogonal to what the article discusses, and could be something that's already happening. The frequency skewing is to keep the Vcore stable after it has been determined what it should be set to. Temperature changes slowly enough that adjustments can be made in a much less tight control loop (and the latencies for changing voltage will be worse than changing frequency anyway)

 

[coercitiv]

The frequency skewing only kicks in when there's a current transient due to idle execution units turning on due to a change in instruction mix. The 0.2% figure is a measured worst case induced by one benchmark. The 7% would just be hit over short durations normally.

Average CPU clock was 3392Mhz for a 3400Mhz nominal speed in the worst case scenario. They gave up 8Mhz for 9% improved Vmin, leading to around 18% power savings.

 

[Exophase]

Average CPU clock was 3392Mhz for a 3400Mhz nominal speed in the worst case scenario. They gave up 8Mhz for 9% improved Vmin, leading to around 18% power savings.

Yeah, about a 0.2% degradation. Which is very small.

But the 9% lower Vmin/18% lower power consumption figure was for operation at 4GHz, while the benchmarks were ran at 3.4GHz. At that frequency the power savings were a little under 14%, according to the graph.

At higher frequencies they'll have higher current swings and will probably droop more. 3.4GHz may have been a good sweet spot for their figures.

 

[coercitiv]

But the 9% lower Vmin/18% lower power consumption figure was for operation at 4GHz, while the benchmarks were ran at 3.4GHz. At that frequency the power savings were a little under 14%, according to the graph.

My bad, they didn't specify Vmin improvement for 3.4Ghz and I had already read about the 9% improvement in the previous paragraphs.

Wonder if this could have a significant impact on their mobile line, leading to easier to sustain turbo speeds. (spanning from 3 to 3.6Ghz)

 

[Phynaz]

Just goes to show there are some very, very smart people working on this stuff.

 

[escrow4]

LOOK LOOK, AMD is now only 5yrs behind Intel, instead of a decade. By the time Steamroller is released, and assuming AMD hasn't fallen apart and given up on CPUs, Intel will be on what, 10 or 14nm?


Thread crapping and trolling are not allowed here.
Markfw900

 

[pw257008]

LOOK LOOK, AMD is now only 5yrs behind Intel, instead of a decade. By the time Steamroller is released, and assuming AMD hasn't fallen apart and given up on CPUs, Intel will be on what, 10 or 14nm?

Kaveri is Steamroller

 

[Arkadrel]

The techniques described will first appear in AMD’s Steamroller based platforms, but are expected to roll out across other IP blocks potentially including GPUs, ARM cores, and the Jaguar core.

Guess that means theres still room for more improvements even on something like Beema/Mullins even if they where to make a "new" again 28nm part.

Even if its only 6-7% or so more, from this.

Second, this technique could be applied to AMD’s discrete and integrated GPUs, although it is hard to say how big the benefits would be for GPUs.

Hmmm could be used on GPUs as well to make them more power effecient.
Cool.

 

[DrMrLordX]

At higher frequencies they'll have higher current swings and will probably droop more. 3.4GHz may have been a good sweet spot for their figures.

Hmm, one wonders if that's why the A8-7600 wound up having such good perf/watt numbers vs. the other members of the Kaveri lineup. In 45w TDP mode, its base clock speed is 3.3 ghz and it turbos up to 3.7 ghz.

 

[rainy]

In 45w TDP mode, its base clock speed is 3.3 ghz and it turbos up to 3.7 ghz.

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

 

[ShintaiDK]

Hmm, one wonders if that's why the A8-7600 wound up having such good perf/watt numbers vs. the other members of the Kaveri lineup. In 45w TDP mode, its base clock speed is 3.3 ghz and it turbos up to 3.7 ghz.

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

 

[Enigmoid]

My bad, they didn't specify Vmin improvement for 3.4Ghz and I had already read about the 9% improvement in the previous paragraphs.

Wonder if this could have a significant impact on their mobile line, leading to easier to sustain turbo speeds. (spanning from 3 to 3.6Ghz)

The problem is that voltage drop will only become a significant problem when significantly loaded which 1 core turbo is less likely to do. It will help though to a lesser extent.

Kaveri Mobile has very aggressive turbo. On the other thread is looks like there is a 19W SKU with 2.0/3.2 speeds. There is no way that these turbo speeds will be maintained when the 45W chips have 3.1/3.3. Adaptive clocking will help though I'm unsure of the amount.

 

[coercitiv]

There is no way that these turbo speeds will be maintained when the 45W chips have 3.1/3.3. Adaptive clocking will help though I'm unsure of the amount.

If this source is right then we might see even higher turbo speeds in the mobile lineup. However, I'd rather wait for a more official release for this data.

 

[Enigmoid]

If this source is right then we might see even higher turbo speeds in the mobile lineup. However, I'd rather wait for a more official release for this data.

Yep, It looks like AMD is going for a very aggressive turbo. This is pretty much the best they can do with their lower efficiency chips. Some of those 19W and 35W chips have almost identical speeds (7500 and 7200P); there is no way the 19W chips will maintain turbo and perform the same as their 35W brethren.

 

[rainy]

Kaveri Mobile has very aggressive turbo. On the other thread is looks like there is a 19W SKU with 2.0/3.2 speeds. .

Is that really so aggressive?
Richland 19W TDP part (A8-5545M) have 1.7/2.7 GHz clocks which means that turbo is quite similar (about 60 percent) to the A10-7300.

Btw, somehow I have a strange feeling that you don't want to see AMD more competitive in mobile market.

 

[AtenRa]

I'm expecting the A10-7500 to be very close to Core i5 4200U in CPU performance. It will be faster in MT loads and a little slower in ST.
It will also be faster in Gaming and OpenCL than the Core i5.

 

[ChronoReverse]

I'm expecting the A10-7500 to be very close to Core i5 4200U in CPU performance. It will be faster in MT loads and a little slower in ST.
It will also be faster in Gaming and OpenCL than the Core i5.

What about power usage? Should I look forward to a notebook with it?

 

[AtenRa]

What about power usage? Should I look forward to a notebook with it?

According to wccf A10-7500 is 19W TDP. It depends on the laptop specifications and quality. Price and availability will be crucial as always. But as far as performance, those Kaveris should be very competitive.

 

[Enigmoid]

Is that really so aggressive?
Richland 19W TDP part (A8-5545M) have 1.7/2.7 GHz clocks which means that turbo is quite similar (about 60 percent) to the A10-7300.

Btw, somehow I have a strange feeling that you don't want to see AMD more competitive in mobile market.

I'm a realist and sick of people hyping the next big "AMD" (or intel BTW-intel WILL delay their future products) product. Heard it about richland, heard it about kaveri (supposed to be 30% faster IPC on some of AMD's slides). I will believe when they deliver. Not saying its impossible but AMD on mobile has been nothing but disappointments and they need to show that they care. Their gpu drivers are horrible on mobile (this is extremely important because a lot of OEM's lock drivers down and don't release updates; you are stuck with a gpu/igp that doesn't work a lot of times--my brother's laptop is like this), still no hybrid crossfire fix. Where is Bonaire on mobile? (they have it for firepro but not consumer chips?) Where is mobile kaveri? Kabini was a half-baked product, missing turbo and proper power gating (did not downclock when idle). Beema/Mullins fixes that but it should not have had to been fixed in the first place.

I would like to see them competitive but the mobile market, especially the Ultralight market is all about efficiency and they simply can't compete with intel.

And richland ULV wasn't able to keep the turbo up.

I'm expecting the A10-7500 to be very close to Core i5 4200U in CPU performance. It will be faster in MT loads and a little slower in ST.
It will also be faster in Gaming and OpenCL than the Core i5.

The i5-4200U is about 5% slower than the a10-5750m. Sorry but a 19W chip is not getting anywhere close to that. I expect it to be quite a bit slower in ST workloads and a little slower in MT loads. It will be faster at gaming and much faster in OpenCL.

Binning will allow for more efficient chips on the ULV end but it can only go so far. You are not going to get 85% of the performance of the a8-7600 (45W) on less than half the power without some caveats. There have also been multiple discussions about the 45W kaveri using more power than the 45W richland under real-world load making efficiency much more of a wash.

 

[AtenRa]

The i5-4200U is about 5% slower than the a10-5750m. Sorry but a 19W chip is not getting anywhere close to that. I expect it to be quite a bit slower in ST workloads and a little slower in MT loads. It will be faster at gaming and much faster in OpenCL.

Binning will allow for more efficient chips on the ULV end but it can only go so far. You are not going to get 85% of the performance of the a8-7600 (45W) on less than half the power without some caveats. There have also been multiple discussions about the 45W kaveri using more power than the 45W richland under real-world load making efficiency much more of a wash.

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.