Taking Radeon ROCm 2.0 OpenCL For A Benchmarking Test Drive

Under the clpeak integer compute test, the RX Vega 64 came in just ahead of the GTX 1080.

The OpenCL-supported FAHBench for Folding@Home failed to run on the Radeon cards with ROCm.

Under LuxMark, the Luxball HDR scene worked with ROCm 2.0 while the microphone and hotel scenes still hung during the CL kernel compilation stage. For the scene that did work, the RX Vega performance was very strong and in fact just ahead of the RTX 2080... LuxMark has long performed very well on AMD Radeon GPUs, well, when not running into problems like the kernel compilation hang.

The SHOC benchmark was running fine with the ROCm 2.0 OpenCL stack and the red performance was good with the RX 580 beating out the GTX 1070 and the RX Vega GPUs surrounding the GTX 1080 Ti.

In cases where ROCm 2.0 is working out nicely, the Radeon OpenCL performance is very strong against the NVIDIA GPUs with their Linux OpenCL driver. Unfortunately though several of the interesting Linux OpenCL test cases were still not behaving well with ROCm 2.0, but hopefully that will improve in subsequent ROCm 2.x releases as we enter 2019. I'm also working on evaluating ROCm 2.0 TensorFlow performance hopefully in the next few days.

With ROCm 2.0 out the door, the AMDKFD kernel driver in mainline becoming more mature, Vega 7nm GPU support appearing already in order, and other open-source driver improvements, 2019 could certainly be an interesting year for Radeon Linux GPU compute... At least from a software perspective; it will be interesting to see what AMD's long-awaited Navi hardware (and any upcoming Vega updates) will be able to deliver in better competing against NVIDIA's latest GPUs. In 2019 we are also expected to see OpenCL-Next where NVIDIA is expected to support that revision as well after not fully supporting OpenCL 2.x over SVM requirements, which are expected to be made optional in this next specification release.