AMDGPU Performance Tests With New WattMan-Like Settings, Power Capping

[bridgman]

I am still wait for a free OpenCl 2.1 library implementation...
Until now only for amdgpu-pro
Kind of sad..

Something like this ?

https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime

I believe the level of OpenCL support is the same as amdgpu-pro's... 2.0 for language and 1.2 for runtime.

[bridgman]

Interesting... and nice round of testing.

Michael , could you take a look at the third-last RX 580 graph from the end (Meta performance per watt) ? It looks more like "meta performance" at each power level rather than "meta performance per watt". The Vega graph looks more consistent with the per-game performance per watt graphs.

[Michael]

Interesting... and nice round of testing.

Michael , could you take a look at the third-last RX 580 graph from the end (Meta performance per watt) ? It looks more like "meta performance" at each power level rather than "meta performance per watt". The Vega graph looks more consistent with the per-game performance per watt graphs.

I would ignore the meta results. At least on the RX Vega 64, I ended up dropping Deus Ex: Mankind Divided as it seemed to be unstable in some of the scenarios, so dropped it after a sub-set of the tests were run, but that would have left the original test runs in a slightly different state/value for their meta value that was computed at test-time.

[tuxd3v]

Something like this ?

https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime

I believe the level of OpenCL support is the same as amdgpu-pro's... 2.0 for language and 1.2 for runtime.

I believe too, but I have a kavery APU( AMD A10-7800 Radeon R7, 12 Compute Cores 4C+8G )

It fails to build the amdgpu 1.8-151, module for kernels,

4.9.0-6-amd64
4.16.0-0.bpo.1-amd64
4.16.0-0.bpo.2-amd64

I read the supported hardware, and its not clear the level of support for Kavery,
for what is presented it should work with internal R7, but not with dGPU..it seems limited.

[darkbasic]

Something like this ?

https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime

I believe the level of OpenCL support is the same as amdgpu-pro's... 2.0 for language and 1.2 for runtime.

Very nice but... does it work with my HD7950? No, it doesn't.

[Guest]

One of the things I've always liked about Linux was getting all the features we need with none of the excessive bloat that the Windows equivalent gets. Sure, a GUI would be appreciated, but it shouldn't take 300MB+ of crap to get that.

I hope you enjoy having your extra free 300MB (worth $0.035 on a cheap 1TB SSD) and not having a proper GUI for these kinds of things

[humbug]

One of the things I've always liked about Linux was getting all the features we need with none of the excessive bloat that the Windows equivalent gets. Sure, a GUI would be appreciated, but it shouldn't take 300MB+ of crap to get that.

I don't know how many megabytes it would be but A GUI for this functionality is very important for user friendliness lke on windows. Maybe AMD can offer it as a separate standalone download.

[pete910]

Out of interest, Is wattman done in QT @bridgman?

Could it not be made cross platform? if AMD allowed obviously.

[Shevchen]

Michael as some people already suggested, the best way to test the Wattman-performance is to first undervolt and then overclock.

A (short) tutorial from the results I got from extensive tweaking:

Under my Vega 64, the first thing I did was to undervolt the card well below 1100mV on the core (a good starting value is 1070 mV) and watch how far I can push the frequency with it. If the core frequency went up and down depending on load, I increased the power target until the frequency was stable. Once I achieved a stable frequency, I tried to tweak the power draw by lowering the Vcore until the bench crashed again. (My chip is a potato, but I was able to push my card to 1597 MHz on the core with 1055 mV at +27% Power Target)

Don't worry about the Power Target too much, just increase it until everything is stable. My card draws well below 200W of power even with the +27% PT.
The idea behind it is to reduce power draw to reduce leakage and heat to give the HBM2 more headroom. If you end up with ~73°C on the card on full load, you are already in the good zone to try to overclock the HBM2 now. My card was able to hit 1065 MHz before artifacting, but other peeps got close to 1200 MHz. Here you just increase the frequency and see if it stays stable. If it drops to 925/800 MHz (Vega 64/56) you need to either tweak the card more to have it run cooler, or your PT it slightly too low.

In the end, I got ~13% more performance out of my card compared to stock (Superposition bench for adaptation and stress test with a game of your choice that pushes your card to the limit) and it consumes <200W of power under full load - while having a mediocre chip from the silicon lottery.

[schmidtbag]

I don't know how many megabytes it would be but A GUI for this functionality is very important for user friendliness lke on windows. Maybe AMD can offer it as a separate standalone download.

I agree, but using a Qt or GTK interface ought to keep it very functional while remaining very lean.