AMD Southern Islands vs. NVIDIA Fermi/Kepler On Gallium3D

[Veerappan]

even when locked to their lowest speeds the top-end Nvidia cards still outperform the Radeons.

Although Im very curious whether Michael was using the auto-generated firmware created by the Nouveau driver or if he actually extracted the binary Nvidia driver firmware for use in the experiment.

Hint: My Radeon 7850 also defaults to a low default clock speed. The high-performance clock specified in the VBIOS is 860Mhz, but the default clock speed on the 3.9 kernel is ~160-180Mhz. The memory is similarly slowed down from the high-performance speeds.

We'll see what happens when some tests of 3.11 kernels are done with radeon.dpm=1. I have a feeling that the SI cards will get a bit faster at that time.

[curaga]

I find it rather sad that you compare a newer Radeon driver, with an older Nouveau driver...

Hint: they are exactly as new, from the very same mesa version.

[gotwig]

Sad

I find it rather sad that you compare a newer Radeon driver, with an older Nouveau driver...

[Calinou]

Micheal, thanks for doing the benchmark--I remember the requests for this.

But, could someone run this benchmark with an HD5xxx/HD6xxx in a similar system (quad-core i7 4770k with HT @ 3.5 GHz/16 GB RAM, 1920x1080 display, SSD), or an intel gpu?
At least I think that's a quad-core with hyperthreading...Micheal, is that correct, or did you get an octocore?

Very roughly, it looks like the 7850 is competing with the GTX 550, and the the 7950 with the GTX 680.
Reaction Quake runs better on nouveau.

The 4770K is a quad core with HT, yes. The only octacore Intel CPUs are Xeons, and they cost at least 3000?.

[AnonymousCoward]

Nah, it's just Gallium3D for SI not being in the greatest shape (compared to r600)

Michael didn't mention what power management option he was using either. I assume he was using the old style power management, but he didn't say if he forced the high GPU profile or not. He also didn't specify which shader compiler was used. I believe the LLVM compiler is the default now, but occassionally the old compiler is faster (both are slower than R600sb of course, although that is unsupported with SI)

[Ancurio]

even when locked to their lowest speeds the top-end Nvidia cards still outperform the Radeons.

Nah, it's just Gallium3D for SI not being in the greatest shape (compared to r600)

[RussianNeuroMancer]

even when locked to their lowest speeds the top-end Nvidia cards still outperform the Radeons.

You need my help to compare prices of tested GPU or you can search for this info on your own?

[andy_v]

Michael, can you please look into whether the HD 7000 series cards use the lowest power profile by default? If so, can you change the profile to high before doing the tests?

[Sonadow]

even when locked to their lowest speeds the top-end Nvidia cards still outperform the Radeons.

Although Im very curious whether Michael was using the auto-generated firmware created by the Nouveau driver or if he actually extracted the binary Nvidia driver firmware for use in the experiment.

[Ibidem]

At last!

Micheal, thanks for doing the benchmark--I remember the requests for this.

But, could someone run this benchmark with an HD5xxx/HD6xxx in a similar system (quad-core i7 4770k with HT @ 3.5 GHz/16 GB RAM, 1920x1080 display, SSD), or an intel gpu?
At least I think that's a quad-core with hyperthreading...Micheal, is that correct, or did you get an octocore?

Very roughly, it looks like the 7850 is competing with the GTX 550, and the the 7950 with the GTX 680.
Reaction Quake runs better on nouveau.