For years, Nvidia has ruled AI with its pricey GPUs thanks to CUDA, the software glue that keeps everyone locked in tighter than Apple's App store. AMD has been sulking on the sidelines, insisting its ROCm tools are ready for prime time if only the world would notice.

At its Advancing AI event, AMD promised that ROCm 7.0 would deliver shinier frameworks, algorithms and enterprise features. Phoronix spotted the rocm-7.0.0 tags on GitHub for HIP, AOMP and the usual library suspects, meaning launch is only weeks away.

The biggest boast is a 3.5x performance boost in AI workloads compared with ROCm 6, along with “accelerated training performance.”

AMD even claimed its Instinct MI355X GPU outpaced Nvidia’s Blackwell B200 by 30 per cent in FP8 throughput for DeepSeek R1. Naturally, this was shown in AMD’s slide deck, so grab a pinch of salt the size of a Zen 5 CCD.

ROCm 7 will add support for the MI350 series, cluster management tools, scaling features and enterprise fluff to make CIOs feel important. The real question is whether developers who have already sold their souls to CUDA will bother to switch.

No official release date has been pinned down, but with GitHub tags in place and marketing engines spinning, expect AMD to drop ROCm 7 in the coming weeks just in time to remind everyone it still wants to play in AI.