LLVMpipe & OpenSWR OpenGL Riding Off Threadripper

[phoronix]

Phoronix: LLVMpipe & OpenSWR OpenGL Riding Off Threadripper

One of the unique test requests coming in as part of our Threadripper on Linux testing is to see how well the LLVMpipe and OpenSWR CPU-based OpenGL implementations within Mesa perform for this 16 core / 32 thread single-socket processor. Here are those results...

http://www.phoronix.com/scan.php?page=news_item&px=LLVMpipe-SWR-Threadripper

[Meteorhead]

Do you happen to have any Knights Landing hardware to race against Epyc, Threadripper, Xeon and Core i9? Would be interesting to see if a 72C/288T beast can make any significant difference?

Anyhow, thanks for the SWR tests lately. Insightful input.

[Michael]

Do you happen to have any Knights Landing hardware to race against Epyc, Threadripper, Xeon and Core i9? Would be interesting to see if a 72C/288T beast can make any significant difference?

Anyhow, thanks for the SWR tests lately. Insightful input.

Nope, I have no KNL hardware, that's why you haven't seen benchmarks on it on Phoronix :P

[muddymind]

Very interesting. I'm sure that with a lot of optimization work a purely software renderer could become very competitive against low-end graphic cards. It reminds me of the good old days when unreal engine had an amazing software renderer.

[devius]

Interesting. Seems to perform worse than the 40 core machine.... who would've guessed? Still the CPU utilization is only about 31% which means that there are possibly ways to optimize software rendering further.

[chithanh]

the CPU utilization is only about 31% which means that there are possibly ways to optimize software rendering further.

In the default build configuration, llvmpipe currently uses a maximum of 16 threads.

[schmidtbag]

I'm honestly a bit surprised how well the HD 4650 performed. Sure, most of these titles aren't that demanding by today's standards, but that GPU aged nicely.

In the default build configuration, llvmpipe currently uses a maximum of 16 threads.

Weird... Well lets hope they change that default. Seems the other software renderers must do the same since they all seem to have roughly the same performance.

[VikingGe]

Weird... Well lets hope they change that default.

They need to improve multicore scaling first, *massively*. llvmpipe doesn't even scale properly to 4 cores.

[bridgman]

Interesting. I couldn't find any common tests between the systems (OpenArena had different resolution, Xonotic had different resolution and quality) but what struck me was the big jump in frame rates for Xeon between Mesa 17 and 17.3 with no corresponding jump for Threadripper. Looks like some llvmpipe changes might have gone in between the versions which were more Intel-specific than intended.

[chithanh]

Looks like some llvmpipe changes might have gone in between the versions which were more Intel-specific than intended.

Looks like AMD needs to step up their internal automated performance and regression testing on Linux. Maybe Robert will let you use those computers for it.