An engineer from AMD's Nod.ai acquisition published a feedback post on AMD's GitHub ROCm support page, seeking guidance on additional GPU support the ROCm community would like to be added to AMD's list of supported ROCm devices.
The engineer put together a poll consisting of various GPU lineups and their associated architectures: the RX 7000 series (RDNA 3), the RX 6000 series (RDNA 2), Phoenix mobile APUs (RDNA 3), Strix / Strix Halo mobile APUs (RDNA 3.5), and RDNA 1 (i.e., the RX 5000 series). The last category was labeled as GCN and prior architectures.
So far, the RX 6000 series RDNA 2 GPUs have won the leading category, followed by Strix / Strix Halo and the RX 7000 series. RDNA 1 has received the fewest votes, while "GCN and prior architectures" take fourth place.
The comments section also focused on what GPUs AMD should support with ROCm. The highest-upvoted comment recommended supporting any GPU with 16GB or more VRAM, including integrated GPUs that can occupy 16GB or more memory through system DRAM. The second-most highly upvoted comment recommended supporting ROCm for all relatively recent GPUs from the last five to six years.
The AMD ROCm open-source software platform currently supports all modern AMD Instinct GPUs, particularly those based on iterations of AMD's CDNA architecture. Likewise, all of AMD's modern workstation graphics cards support ROCm on Linux.
Where support gets touchy is Windows support and support for Radeon-based products. The only Radeon GPUs supported on Linux are the RX 7900 series graphics cards and the Radeon VII. On Windows, most of the RX 6000 series and all of the RX 7000 series desktop GPUs support ROCm, but the RX 6700 - RX 6750 XT don't support the HIP SDK, only Runtime. Similarly, the Radeon Pro W7900, down to the W6600, is supported, but the W6600 lacks HIP SDK support, and the Pro VII and W5500 aren't supported on Windows. No mobile GPUs support ROCm, according to AMD's device support list.
ROCm is an open-source software platform developed by AMD to aid programmers in running HPC and AI applications on AMD GPUs. It is essentially AMD's counterpart to Nvidia's CUDA architecture, but unlike Nvidia, it does not support all AMD GPUs dating back to the mid-2000s.
