Testing Out The Configurable TDP On AMD's Kaveri

[edgar_wibeau]

I also can't find numbers about the resulting CPU frequencies for the cTdp settings. That would be very helpful, too.

Anyway, thanks for the first cTDP related benchmarks I can find in the net!

[AJSB]

Michael, can we have actual chart(s) about the real power drain when cTDP is set to 45W ?
Because i know that one thing is what we set it for and another is what it is really spending....and i know from A8-7600 reviews.

AMD says that other Kaveris besides A8-7600 will be less eficient with cTDP set, so, i was curious about actual values of a A10-7850K set to 45W.

[Spittie]

You do realize that TDP is a measure of heat production not energy draw right? Hence it expanding to: Thermal Design Power...

Yes, but it's obviously going to draw less current if it has to release less heat
I know it's not exactly the best way to compare stuff, but I think that it still show how well Kaveri does at the lower end.

@Everyone else
Yes, FPS/W.
I realized that I've got it flipped after 5 minutes (stupid 5 minutes limit...), and my second post was put into moderation for whatever reason.

[Luke_Wolf]

You do realise that the laws of thermodynamics still applies? If it's producing 45W of heat, it's at least using 45W of electricity.

Not quite because the TDP defines an approximate of the maximum amount of heat that the processor is allowed to produce, and has no bearing to how much heat is produced during any particular workload, only that the heat produced is (supposed to be) < TDP, now how this is measured though varies between vendors as guido said. AMD pushes their processor to the wall when determining this, Intel not so hard.

[wizard69]

Interesting things are afoot in AMD land.

The current model only implements two of the four possible memory controllers. So it looks like AMD has the intention to address the RAM bandwidth issue one way or another. The another being GDDR graphics RAM.

Yes, it's the RAM-bound GPU that makes the GPU look 'unaffected' from what I can tell. There was a discussion about this a couple weeks ago in another article which lead Michael to then go and do another article based on Kaveria (sorry, no link. I just use the RSS feeds) a couple days ago about this which pretty much backed up the theories (which was originally about how bound to RAM the onboard GPU's from even the early days actually were).

This has always I been an issue with APUs. In fact it was an issue when GPUs wgphere integrated into external chip sets. Nothing beats bandwidth for a GPU.

It's why I've always bought the highest compatible RAM for my HTPC's and low-end machines. No point burning power that's unused because the GPU was throttled. But now we can reduce the power usgae (with proof yet to come on ACTUAL power consumption) and still get damn fine GPU-based performance. Good times all-round I say, especially in light of games not really requiring much in the way CPU performance.

This new APU of AMDs is a bigger leap forward than a lot of people want to admit to. Frankly these days there are many applications where the CPU not the most important element of the machine, it isn't just gaming.

[guido12]

Cool! Thanks for the 45 W and 65 W TDP numbers! I assume they'd be in the same ballpark as the A8-7600.

As for TDP, it's my understanding that it is determined by measuring the average (or peak minus outlier spikes) power consumption during a particular type of prolonged stressful CPU load and case temperature. For Intel, their TDP measurements are done doing some "typical" software loads while AMD uses more stressful loads (ie. benchmarking software) so AMD TDPs and Intel TDPs can't exactly be compared directly. I guess with integrated GPUs they also add GPU loads.

[gigaplex]

You do realize that TDP is a measure of heat production not energy draw right? Hence it expanding to: Thermal Design Power...

You do realise that the laws of thermodynamics still applies? If it's producing 45W of heat, it's at least using 45W of electricity.

[Adarion]

This is one of the features I found really interesting. New, tech-stuffish and definitely interesting. While you won't enable it on raw compiling, video rendering or excessive gaming, you might throttle it down nicely for HTPCs, surf stations and so on. You could even buy them in 10 packs with a rebate and then put them in different boxes for different purposes. Just configure one to be your little HTPC with ASIC aided video decoding and one as you big boom machine for Gentoo compiling or similar things.

The mainboard looks also interesting though I'd rather go with the A88X-Plus cause it has 3 classic legacy old school PCI slots (for my bazillion of PCI cards I still have and use), as well as lots of SATA and USB headers.

[Luke_Wolf]

Interesting results, Kaveri really does good when paired with a low TDP.
It's also showed by the A8-7600, which in most benchmarks is always just 10/15% slower than the A10-7850K, while having a lower TDP and a lower clock. (Too bad that AMD didn't send you one of those too, Tje A8 should be ~$70 cheaper than the A10 when it hit the market, so a comparision under Linux would have been interesting).

It really shows that Kaveri was optimized to work with low TDP. Or maybe it's just because it's using 28nm SHP, who knows.


It's also interesting to see the Watt per FPS performance (results rounded, I've assumed that auto is 95W):
Xonotic
45W - 2.95 W/fps
50W - 2.8 W/fps
55W - 2.63 W/fps
60W - 2.43 W/fps
65W - 2.3 W/fps
95W - 1.71 W/fps

x264
45W - 1.53 W/fps
50W - 1.48 W/fps
55W - 1.4 W/fps
60W - 1.31 W/fps
65W - 1.26 W/fps
95W - 0.97 W/fps

You do realize that TDP is a measure of heat production not energy draw right? Hence it expanding to: Thermal Design Power...

[stiiixy]

From the benchmarks, CPU performance seems to be much more constrained by the TDP than GPU. So either the GPU is extremelly efficient or it is bottlenecked by something else, which may very well be the CPU itself. Maybe overclocking the CPU will render further FPS gains in graphics benchmarks.

Yes, it's the RAM-bound GPU that makes the GPU look 'unaffected' from what I can tell. There was a discussion about this a couple weeks ago in another article which lead Michael to then go and do another article based on Kaveria (sorry, no link. I just use the RSS feeds) a couple days ago about this which pretty much backed up the theories (which was originally about how bound to RAM the onboard GPU's from even the early days actually were).

It's why I've always bought the highest compatible RAM for my HTPC's and low-end machines. No point burning power that's unused because the GPU was throttled. But now we can reduce the power usgae (with proof yet to come on ACTUAL power consumption) and still get damn fine GPU-based performance. Good times all-round I say, especially in light of games not really requiring much in the way CPU performance.