AMD ROCm 6.4 Arrives: Performance Boosts but RDNA 4 Support Absent
AMD just released ROCm 6.4, the latest version of their software platform. ROCm is designed to tap the power of AMD GPUs for high-end computing. AI, machine learning, and high-performance computing are ROCm's specialties. This latest version brings many welcomed changes and greater compatibility. However, one glaring omission has programmers stumped: added support for RDNA 4 GPUs.
RDNA 4 Support Missing in ROCm 6.4
RDNA 4 is AMD's new graphics architecture. It is the brains behind recently launched Radeon RX 9000 series GPUs. These graphics cards are replete with architectural additions. These additions are specifically targeted to enhance performance in workloads that ROCm targets. RDNA 4 doubles the FP16 per cycle. It offers an eight-fold increase in INT4 performance with sparsity. FP8 support provides another eight-fold improvement over RDNA 3 in FP16 work. On paper, RDNA 4 is powerful.
The problem is the lack of ROCm official support. Without it, all that potential muscle remains idle. It's a high-performance sports car without an ignition key. Navi 48 rumors, the GPU used in the RX 9000 series, did surface within ROCm patch notes last year. This only adds to sentiments of hope and now disappointment.
ROCm 6.4: What's New
ROCm 6.4 does introduce good updates. These include:
- CPX Mode with NPS4 Memory: CPX memory mode support for better performance in certain workloads.
- PyTorch Power-Up: Support for the latest PyTorch frameworks, versions 2.6 and 2.5. This places developers at the leading edge of AI development.
- VP9 Video Decoding: VP9 video codec support through rocDecode/rocPyDecode. This improves media processing capabilities.
- Enhanced Profiling Tools: Reworks on the ROCm Compute Profiler. This enables developers to optimize code for AMD GPUs.
- Boosted OS and Hardware Support: Support for Oracle Linux 9 and Radeon PRO W7800 GPUs. This extends the overall ecosystem.
The CUDA Challenge and Inconsistent Support
Despite these improvements, AMD's failure to take Nvidia seriously in professional GPU computing remains a key issue. Nvidia's CUDA platform is the de facto standard. One reason for this is its consistently better and timely hardware support. AMD's ROCm is open-source and promising, but it plays catch-up.
ROCm support on Windows is being extended to most of the RDNA 2 and RDNA 3 families. This was previously announced in 2022. However, major exceptions exist like the RX 7650 GRE and RX 7900 GRE. HIP SDK support does not exist for the RX 6600 to RX 6750 XT segment. Linux compatibility is even more constrained. It is available on a very small set of Radeon GPUs.
Strix Halo Support: A Positive Sign
ROCm 6.4 does support AMD's Strix Halo APUs. These performance-oriented chips have up to 128GB memory. They are now attractive options for mobile AI and HPC application development. This is a welcome step forward. It provides a portable platform for ROCm workloads.
Developer Patience Tested
AMD tends to release ROCm updates monthly. RDNA 4 support might appear in the next generation. However, the repeated lag in compatibility raises a question. How much longer will developers wait? Nvidia offers a more predictable and stable environment. For developers needing early access to the latest GPU architecture, CUDA and Nvidia's strong support are appealing. The clock is ticking for AMD to close this gap. ROCm needs to keep pace with their cutting-edge hardware advances.
