- Sources: AMD Advancing AI 2026 announcement, AMD Instinct MI455X product page, AMD and Anthropic partnership, Chips and Cheese analysis
- Summary: AMD launched the Instinct MI400 Series, 6th Gen EPYC "Venice" CPUs, and the Helios rack-scale system at its Advancing AI event on 2026-07-23. AMD's product page gives the MI455X as CDNA5 on TSMC 2nm and 3nm, 320 billion transistors, 256 work group processors at a 2.4 GHz peak engine clock, 432 GB of HBM4 across 12 stacks at 23.3 TB/s, 192 MB of L2, and 40.3 PFLOPs peak OCP MXFP4, with a launch date of 2026-07-23. A Helios rack pairs 72 MI455X GPUs with 18 EPYC Venice CPUs over AMD Pensando networking and the ROCm software stack. AMD states Helios delivers up to 30% more inference tokens per dollar than the leading competing solution, a vendor figure with no published method. AMD names OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr, and Cirrascale as adopters, and says OpenAI expects Helios online beginning in the fourth quarter of 2026. Anthropic and AMD announced separately on 2026-07-22 a deployment of up to 2 gigawatts of MI450 Series GPUs in Helios racks with the first gigawatt starting in the first half of 2027, an AMD commitment to a strategic equity investment of up to $5 billion in Anthropic, and a multi-year collaboration using Claude to optimize workloads for Instinct GPUs and accelerate ROCm development. Chips and Cheese reports each work group processor is now four dual-issue Wave32 SIMD32 units in place of CDNA4's four single-issue Wave64 SIMD16 units, so the large gains are on FP4 and FP8 while the rest comes from the wider SIMD.
- Why it matters: A second source of frontier-scale training and inference hardware now has a rack-scale product and named gigawatt-scale customers, and the path to using it runs through ROCm rather than CUDA.
- Follow-up: Watch for independent Helios benchmarks against the tokens-per-dollar claim, the first OpenAI deployments in the fourth quarter, and whether the Claude-for-ROCm collaboration publishes results.
send feedback on this story