AMD Talks Up Imminent ROCm 6.3 With Big Performance Gains, New Features

Either due to a mistimed blog post or other factors, a big feature article is out talking up the new ROCm 6.3 features... But the updated ROCm 6.3 open-source GPU compute software doesn't appear to actually be released yet at all their usual sources. In any event there are new features and big performance gains being talked up for ROCm 6.3.

A blog post this evening on community.amd.com covers all the ROCm 6.3 highlights. The only problem is that ROCm 6.3 isn't actually out yet. For the many ROCm GitHub repositories I am subscribed to there hasn't been any ROCm 6.3 tags/releases yet. The ROCm documentation portal is still pointing to ROCm 6.2.4 as the latest. As of writing the repo.radeon.com/rocm/ repository / package archives for the various Linux distributions do not yet have ROCm 6.3 yet either.

Whatever the case, it appears the ROCm 6.3.0 release is imminent given the blog post this evening. And with ROCm 6.3 there is going to be a number of performance optimizations as well as new features. Some of the ROCm 6.3 highlights from the AMD blog post today note:

- SGLang is a new runtime supported by ROCm 6.3 that is "purpose-built for optimizing inference of cutting-edge generative models such as LLMs and VLMs on AMD Instinct GPUs." With SGLang AMD is talking up to 6x higher throughput performance on LLM inferencing, ease of use with Python integration, and more.

- FlashAttention-2 optimized for ROCm 6.3 is talking up to 3x speed-ups and other benefits.

- The new AMD Fortran compiler is included in ROCm 6.3 with direct GPU offloading, backward compatibility, and other benefits.

- New multi-node FFT within rocFFT for more powerful HPC workflows.

- AV1 video codec support for media processing within rocDecode and rocPyDecode.

- GPU-accelerated JPEG decoding with the rocJPEG library.

- Better audio augmentation with the rocAL library.

- ROCm 6.2's Omnitrace has been re-branded as the ROCm System Profiler and Omniperf is now known as the ROCm Compute Profiler.

ROCm 6.3 sounds like quite an exciting release... I look forward to its actual debut and for hopefully finding the time for some fresh testing/benchmarking.

This end-of-year release of ROCm 6.3 is also a bit unusual as I was sort of figuring AMD would be releasing ROCm 7.0 around this time. ROCm 6.0 came in December 2023, ROCm 5.0 in early 2022, ROCm 4.0 in December 2020, ROCm 3.0 in November 2019, and ROCm 2.0 in December 2018. ROCm is roughly on a one year cadence between major X.0 releases and they typically have landed around the end of each calendar year... So now seeing this big "ROCm 6.3" release coming at late November or so seems to throw off that cadence. ROCm 6.3 is on the heavier side so I would be surprised if there is suddenly a big ROCm 7.0 release in the next month or two.

During the summer was also talk of the new AMD Unified AI Software Stack coming around the end of the year. But given the timing of ROCm 6.3 and without signs of ROCm 7.0 leaves me curious if that's delayed into next year with ROCm 7.0 or what's going on or if this AMD Unified AI Software Stack will simply layer over the top of the current ROCm software and otherwise be independent.

In any event, this AMD blog post is out tonight going over all the exciting ROCm 6.3 features while we look forward to seeing what more is on the horizon as we roll into 2025.