Taking Radeon ROCm 2.0 OpenCL For A Benchmarking Test Drive

Last week AMD officially released ROCm 2.0 as the newest major release of the Radeon Open Compute stack. Here are some initial benchmark figures for that Radeon Linux compute component on Polaris and Vega hardware.

ROCm 2.0 was released last week with Vega 7nm (Vega 20) support, a new library (MIvisionX), updated TensorFlow port to ROCm, 48-bit virtual addressing for Vega, more OpenCL 2 work, and an assortment of other changes. For your viewing pleasure today are some initial OpenCL benchmark figures with a Radeon RX 580, RX Vega 56, and RX Vega 64 while more ROCm 2.0 tests will be coming in the new year.

ROCm 2.0 tests were done using AMD's official release binaries on Ubuntu 18.04.1 LTS. The Radeon RX 580, Radeon RX Vega 56, and Radeon RX Vega 64 graphics cards were tested on the red side. On the green side was the NVIDIA 415.23 driver with the GeForce GTX 980 Ti, GTX TITAN X (GM200), GTX 1060, GTX 1070, GTX 1080, GTX 1080 Ti, RTX 2080, RTX 2080 Ti, and TITAN RTX based upon the available and supported cards i had from both vendors.

Tests were done on the Intel Core i9 9900K + ASUS PRIME Z390-A system with 16GB of RAM and 250GB Samsung 970 EVO NVMe solid-state drive for the Ubuntu 18.04.1 LTS installation.

The OpenCL Darktable performance with ROCm 2.0 performance was a bit odd with the RX 580 and RX Vega 56 performing very slowly compared to the RX Vega 64 and the NVIDIA cards. The RX Vega 64 was at least aligned with the NVIDIA OpenCL test cases but was coming in slow -- shy of the GTX 980 Ti performance for this particular program.

With the Parboil scientific program running on OpenCL, the RX Vega cards were competing mostly with the NVIDIA Maxwell era GPUs while the RX 580 was even slower.

ROCm 2.0 was still failing to work correctly for the Rodinia OpenCL test cases.

With the Mixbench OpenCL program we have our first good showing out of the Radeon hardware on ROCm 2.0. In the single precision test, the RX Vega 56 comes out ahead of the GTX 1080 and the RX Vega 64 just ahead of the GTX 1080 Ti and RTX 2080. The RX 580 meanwhile was coming in between the GTX 1060 and GTX 1070 where it's expected.

The basic clpeak benchmark and testing the global memory bandwidth also behaved correctly on ROCm 2.0. The RX Vega performance with its HBM2 is performing strong for the video memory bandwidth against NVIDIA's latest Turing hardware with GDDR6.