AMD MI350X pushes ROCm parity

AMD’s AI push is now less about “can the chip run fast?” and more about “can developers actually live on the software stack?” That’s the real gap Nvidia built around CUDA. AMD’s answer arrived in a more concrete form when it launched the Instinct MI350 series and previewed ROCm 7 on June 12, 2025 — pairing new hardware with a much louder claim that the software is finally ready for mainstream AI work. (amd.com) ### What is the MI350X, exactly? The MI350X is AMD’s fourth-gen CDNA data-center GPU for AI training, inference, and HPC. The headline specs are straightforward: 288GB of HBM3E memory, 8TB/s of memory bandwidth, and support for MXFP6 and MXFP4 low-precision formats that matter for modern inference workloads. AMD positio(amd.com)lide. (amd.com) ### Why do those specs matter? Memory is the first thing to watch. Big models are constrained by how much fits on the accelerator and how quickly weights move in and out. So 288GB and 8TB/s are not vanity numbers — they are AMD’s attempt to make larger models and denser inference practical without heroic system design. AMD also said the broader MI350 family del(amd.com)ncing gains versus the prior generation, though those are vendor benchmarks and should be read that way. (amd.com) ### So is the real story hardware or software? Software — basically. Nvidia’s moat is not just silicon. It’s CUDA, libraries, tooling, and the fact that teams already know how to use it. AMD’s ROCm pitch is that developers can now port and run more of the same frameworks and models without rebuilding everything from scratch. AMD’s current language leans on “eas(amd.com)oader tooling around Kubernetes, operators, and enterprise deployment. (amd.com) ### Has ROCm really hit “95% CUDA parity”? That exact number is not what AMD is officially foregrounding in the material I found. The stronger documented claim is directional, not a single parity score: ROCm 7 adds MI350 support, HIP 7 portability work, distributed inference support, prebuilt containers, and expanded framework coverage. In plain English, AMD is saying the missing p(amd.com)eing the automatic deal-killer. But “parity” still depends on the workload, framework, kernels, and how much custom CUDA code a customer already has. (rocm.docs.amd.com) ### What changed with ROCm 7? ROCm 7 is the software release tied most directly to this push. AMD said the preview brought up to 4x inference and up to 3x training improvement over ROCm 6 in its own tests, plus better support for large-scale inference, model serving stacks, and migration via HIP. That matters because a GPU launch without a matching software step-up is just th(rocm.docs.amd.com)licon are moving together now. (amd.com) ### Are there real customers yet? There are real partnerships, yes — but that is not the same thing as broad production share. AMD has pointed to relationships involving OpenAI, Oracle, Microsoft, Meta, xAI, Cohere, and others in its AI ecosystem messaging. Oracle also announced plans in October 2025 to offer a public A(amd.com) that some of the biggest public capacity headlines are already moving to the next generation. (amd.com) ### What’s still missing? The clean proof point would be a visible wave of customers switching major production AI workloads away from Nvidia at scale because ROCm is good enough and cheaper. That evidence is still thin in public. AMD has a more credible hardware story now, and a much more credible software story than it (amd.com) AI infrastructure. (amd.com) ### Bottom line? MI350X makes AMD’s AI argument easier to take seriously because the hardware is concrete and ROCm is no longer the obvious weak link. But the market is still pricing a future where software progress turns into large, repeatable production wins — and that last step has not been fully demonstrated yet. (amd.com)