AMD released ROCm 6.4, bringing along several improvements, including framework enhancements, broader OS support, and enhancements to several performance and profiling tools. What it continues to lack is enablement for RDNA 4 GPUs, which may discourage developers from shifting to AMD's latest architecture.
When RDNA 4 hit shelves last month, day-one ROCm compatibility was highly anticipated, as was teased by AMD's VP of AI Software. ROCm is AMD's open-source GPU programming platform, rivaling Nvidia's CUDA ecosystem to power HPC and AI applications. ROCm is chiefly focused on AMD's MI Instinct Accelerators and its prosumer Radeon Pro family, but support has slowly been trickling down to consumer-grade Radeon GPUs, just not as fast as CUDA.
RDNA 4 brings several architectural improvements to the table, doubling FP16 operations per cycle, with an x8 increase in INT4 operations with sparsity. Furthermore, with support for FP8, RDNA 4 can dish out an eightfold increase over RDNA3's FP16 capabilities, again with sparsity. Without official ROCm support, these enhancements are effectively dormant. And, ironically enough, the first mention of Navi 48 (the GPU that powers the RX 9070 series) traces back to a ROCm patch last year.
Latest Videos FromTom's Hardware
In any case, AMD introduced several changes with ROCm 6.4, which can be summarized as:
- Support for CPX mode with NPS4 memory mode.
- Support for PyTorch 2.6 and 2.5.
- Support for VP9 with rocDecode/rocPyDecode.
- Several improvements to the ROCm Compute Profiler.
- Support for Oracle Linux 9 and the Radeon PRO W7800 GPU.
Despite recent pushes to challenge the CUDA moat, AMD's hardware support remains perpetually behind Nvidia's. ROCm support for consumer-grade Radeon GPUs on Windows, which began in 2022, now extends across almost all GPUs from the RDNA 2 and RDNA 3 families, excluding HIP SDK support for the RX 6600 to RX 6750 XT range. I said "almost all" as the list lacks the RX 7650 GRE and the RX 7900 GRE. The Linux side is even more dire, compatible with only four Radeon GPUs.
It's not all bad news, as you can now use ROCm on AMD's Strix Halo family of APUs featuring up to 128GB of memory, making them great options for powering AI and HPC workloads on the go. AMD generally releases ROCm updates on a monthly cadence, so we could see RDNA 4 support spring up with the next release. However, this inconsistency could sway developers towards Nvidia, which offers better and more predictable compatibility.
