Radeon ROCm 1.9.1 vs. NVIDIA OpenCL Linux Plus RTX 2080 TensorFlow Benchmarks

In SHOC's OpenCL FFT single-precision test, Vega is running particularly strong with ROCm 1.9 with the RX Vega 64 managing to run in-line with the RTX 2080 for this basic but important test. The RX Vega 56 meanwhile was coming in behind the GTX 1080 Ti.

But Vega's compute potential isn't across the board greatness... With the MD5 hashing, the RX Vega 64 falls back to where it commonly performs of being between the GTX 1080 and GTX 1080 Ti for compute.

With SHOC's texture read bandwidth test, even with the Vega's HBM2 memory the Vega 64 performance was in line with the GTX 1070. The NVIDIA RTX cards with their GDDR6 video memory were easily well past the GDDR5 Pascal cards.

In the cl-mem memory copy test, the RX Vega cards on their latest ROCm release were in line with the GTX 1070/1080 hardware.

LuxMark meanwhile is a test that has long performed quite well on Radeon GPUs. In the case of the Luxball HDR scene, the RX Vega cards managed to perform in line with the RTX 2070/2080. Unfortunately though the two other LuxMark scenes are still hanging during the OpenCL kernel compilation process with the current ROCm release... So Radeon + ROCm is working out well for LuxMark at least when the kernel compilation process doesn't hang, a problem not encountered with the NVIDIA Linux OpenCL driver.