AMD Launches MI455X and Helios, Escalating AI Compute to Rack-Scale

AMD has officially launched the Instinct MI455X GPU and the Helios rack-scale system, marking a shift in AI hardware competition from single chips to complete rack-level system design. With major breakthroughs in memory capacity, process node, and bandwidth, alongside an accelerating software ecosystem, AMD is now positioned to compete directly with NVIDIA at the rack level.

Confirmed

According to the announced specs, the MI455X is built on a 2nm process and delivers approximately 40 PFLOPS of MXFP4 compute. @LysandreJik pointed out that its single-card memory of 432GB HBM4 exceeds the combined total of five H100 GPUs. The architecture achieves over a 3x increase in scale-up bandwidth via 36 UALoE. At the system level, @SumitGup noted that the Helios rack integrates 72 MI455X GPUs and 12 Broadcom chips, delivering 260 TB/s of aggregate rack bandwidth. On the software front, AMD has shortened ROCm's release cadence to 6 weeks; @AnushElangovan and @ryanshrout view this as AMD's push to make its GPU software stack more user-friendly and optimization more automated to close the gap with CUDA. Furthermore, the AMD Advancing AI 2026 event focused entirely on rack-scale infrastructure, ROCm, and broader ecosystem development. Both @AccBalanced and @ryanshrout believe the MI455X breaks the traditional single-chip competition mindset, giving AMD a true rack-level response to NVIDIA.

Unconfirmed

There is a dispute regarding the performance comparison between Helios and NVIDIA's Vera Rubin. @karlfreund cautioned that AMD's Helios uses a "double-wide rack" configuration, whereas NVIDIA's Vera Rubin uses a "single-wide rack," meaning the two cannot be directly compared on a strict one-to-one basis.

2026-07-24 ~ 2026-07-25 · 9 related posts

Full story(4 episodes)→

Primary sources