AMD rolls out next‑gen AI hardware platform with EPYC CPUs and MI455X GPUs

AMD announced a suite of AI‑focused server hardware at the Advancing AI 2026 event. The company unveiled the MI455X accelerator, built on CDNA 5 with 432 GB HBM4 memory and a claimed 34‑fold increase in token throughput per dollar. The MI455X is integrated into the Helios rack‑scale system that houses 72 GPUs, two sixth‑generation EPYC Venice processors per compute tray and Pensando networking, delivering up to 2.9 EFLOPS of low‑precision performance.

The EPYC Venice family, based on Zen 6, was also introduced, offering up to 256 cores, DDR5‑8000 support, PCIe 6.0 and improved power‑management features for AI and HPC workloads. AMD said shipments of Helios systems will begin in Q3 2026, with early deployments expected by OpenAI later that year.

Separately, AMD partnered with Cerebras on a disaggregated AI platform that separates prompt processing onto AMD Helios systems from token generation on Cerebras’s wafer‑scale engine, aiming to boost inference efficiency. In Taiwan, MiTAC Computing Technology demonstrated AI servers built on the new sixth‑generation AMD EPYC processors, highlighting liquid‑cooled and air‑cooled designs for high‑density data‑center deployments.