The Current CPU Driver Usage Difference Between RADV/RadeonSI & NVIDIA
Yesterday I posted some fresh GPU/driver benchmark results for discrete AMD Radeon and NVIDIA GeForce graphics cards. These were some of the most competitive numbers yet we've seen out of the open-source RadeonSI OpenGL and RADV drivers while using the latest Linux 4.15 kernel, especially for the GTX 1060 vs. RX 580 battle. In the comments were requests to see some CPU utilization numbers, including from one of the Radeon Linux developers, so here is a look at how the CPU usage compares.
With having some spare cycles this morning on that Core i7 8700K "Coffee Lake" desktop, I ran a CPU usage comparison with various Linux games when using the Radeon RX 580 (on Linux 4.15 + Mesa 17.4-dev + LLVM 6.0 SVN) vs. the comparable GeForce GTX 1060 (on Linux 4.15 + NVIDIA 390.12) for showing the latest CPU utilization difference for both OpenGL and Vulkan games.
Using the Phoronix Test Suite makes carrying out such a comparison trivially easy and accurately. Simply run MONITOR=cpu.usage phoronix-test-suite benchmark "the desired tests" and let the automation handle the rest.
At 1080p with OpenGL in Dota 2 when RadeonSI has the slight performance advantage, its CPU usage spiked higher than with the NVIDIA driver.
While the Vulkan performance sides with NVIDIA over RADV, but the CPU usage is quite similar.
With F1 2017, the RX 580 on RADV manages to come out ahead of NVIDIA but with a slightly higher average CPU usage although less spikes than NVIDIA.
For those that requested or are interested in the CPU usage metrics between the Radeon and NVIDIA drivers, you can find a lot more data via this OpenBenchmarking.org result file.
The summary is that the Radeon stack led to slightly higher CPU use still than the NVIDIA Linux driver except the NVIDIA proprietary driver did spike higher occasionally. But given all the activity that goes on in Mesa, we'll see where it ends up by the time of Mesa 18.0~18.1.
If you missed yesterday's large Radeon vs. GeForce dGPU comparison with the latest Linux drivers, see NVIDIA GeForce vs. AMD Radeon Linux Gaming Performance At The Start Of 2018.
With having some spare cycles this morning on that Core i7 8700K "Coffee Lake" desktop, I ran a CPU usage comparison with various Linux games when using the Radeon RX 580 (on Linux 4.15 + Mesa 17.4-dev + LLVM 6.0 SVN) vs. the comparable GeForce GTX 1060 (on Linux 4.15 + NVIDIA 390.12) for showing the latest CPU utilization difference for both OpenGL and Vulkan games.
Using the Phoronix Test Suite makes carrying out such a comparison trivially easy and accurately. Simply run MONITOR=cpu.usage phoronix-test-suite benchmark "the desired tests" and let the automation handle the rest.
At 1080p with OpenGL in Dota 2 when RadeonSI has the slight performance advantage, its CPU usage spiked higher than with the NVIDIA driver.
While the Vulkan performance sides with NVIDIA over RADV, but the CPU usage is quite similar.
With F1 2017, the RX 580 on RADV manages to come out ahead of NVIDIA but with a slightly higher average CPU usage although less spikes than NVIDIA.
For those that requested or are interested in the CPU usage metrics between the Radeon and NVIDIA drivers, you can find a lot more data via this OpenBenchmarking.org result file.
The summary is that the Radeon stack led to slightly higher CPU use still than the NVIDIA Linux driver except the NVIDIA proprietary driver did spike higher occasionally. But given all the activity that goes on in Mesa, we'll see where it ends up by the time of Mesa 18.0~18.1.
If you missed yesterday's large Radeon vs. GeForce dGPU comparison with the latest Linux drivers, see NVIDIA GeForce vs. AMD Radeon Linux Gaming Performance At The Start Of 2018.
23 Comments