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.
A few days back I did a similar LLVMpipe/SWR comparison on a 80 thread Intel system so check out those numbers if you are unfamiliar with these CPU-based OpenGL drivers... This testing is done mostly for curiosity about the viability of LLVMpipe/SWR on CPUs with high core counts.
I tested LLVMpipe on Ubuntu 17.04's default stack (Mesa 17.0.7) and then tested LLVMpipe and SWR when built from Mesa 17.3-dev Git paired with LLVM 6.0 SVN.
And for putting the results into perspective, comparing these software numbers to some low-end Radeon hardware... A Radeon RX 460 and even with that being many times faster than LLVMpipe, also a vintage Radeon HD 4650.
An interesting round of benchmarks to say the least.
The LLVMpipe/SWR performance on the Threadripper 1950X is much better than you would find with older desktop CPUs with lower core counts, cache sizes, and those CPUs without AVX/AVX2, but still it's pretty slow.
LLVMpipe is mostly used as a means of driver fallback for debugging along a vendor-neutral code-path or also is being commonly used as a fallback for various Linux desktop environments when no GPU driver is available. OpenSWR is along similar lines while Intel appears to have actual customers using it for some server-side rendering of visualizations and other off-screen rendering tasks.
LLVMpipe and SWR also only support OpenGL 3.3.
More benchmarks via this OpenBenchmarking.org result file. OpenGL is still best kept on the GPU, even as core counts continue to rise and software drivers make use of AVX, etc.
For those wondering about the CPU utilization on Threadripper when using LLVMpipe/SWR, those curiosity numbers via this OpenBenchmarking.org page. Back to the real-world tests now...
A few days back I did a similar LLVMpipe/SWR comparison on a 80 thread Intel system so check out those numbers if you are unfamiliar with these CPU-based OpenGL drivers... This testing is done mostly for curiosity about the viability of LLVMpipe/SWR on CPUs with high core counts.
I tested LLVMpipe on Ubuntu 17.04's default stack (Mesa 17.0.7) and then tested LLVMpipe and SWR when built from Mesa 17.3-dev Git paired with LLVM 6.0 SVN.
An intentionally awkward CPU/GPU combination to say the least...
And for putting the results into perspective, comparing these software numbers to some low-end Radeon hardware... A Radeon RX 460 and even with that being many times faster than LLVMpipe, also a vintage Radeon HD 4650.
An interesting round of benchmarks to say the least.
The LLVMpipe/SWR performance on the Threadripper 1950X is much better than you would find with older desktop CPUs with lower core counts, cache sizes, and those CPUs without AVX/AVX2, but still it's pretty slow.
LLVMpipe is mostly used as a means of driver fallback for debugging along a vendor-neutral code-path or also is being commonly used as a fallback for various Linux desktop environments when no GPU driver is available. OpenSWR is along similar lines while Intel appears to have actual customers using it for some server-side rendering of visualizations and other off-screen rendering tasks.
LLVMpipe and SWR also only support OpenGL 3.3.
More benchmarks via this OpenBenchmarking.org result file. OpenGL is still best kept on the GPU, even as core counts continue to rise and software drivers make use of AVX, etc.
For those wondering about the CPU utilization on Threadripper when using LLVMpipe/SWR, those curiosity numbers via this OpenBenchmarking.org page. Back to the real-world tests now...
20 Comments