AMD MI455X Architecture Breakdown: First Rack-Native GPU with HBM4
ryanshrout · x · 2026-07-24
AMD has introduced the Instinct MI455X GPU, designed specifically for rack-scale AI rather than competing purely at the chip level. The architecture features several major breakthroughs:
- Compute & Bandwidth: Delivers roughly 40 PFLOPS of MXFP4 compute. It increases bandwidth to 3.6 TB/s via 36 UALoE links, allowing 72 GPUs within a rack to act as a single shared-memory system.
- Memory Advantage: Packs 432 GB of HBM4 with 23.3 TB/s bandwidth and grows the L2 cache to 192 MB, heavily optimizing bottlenecks for Mixture-of-Experts models and long-context inference.
- Architecture Overhaul: Transitions from Wave64 to a native Wave32 execution model, increases addressable registers to 1,024 per thread, and introduces a dedicated Tensor Data Mover.
- Advanced Packaging: Utilizes TSMC's 2nm process for 8 compute dies, combined with 3D stacking and CoWoS-L packaging, totaling 320 billion transistors.
Related event: AMD Launches MI455X and Helios, Escalating AI Compute to Rack-Scale(9 posts)→
More from Infra
- Anthropic's ~11 GW contracted compute estimated at ~42% NVIDIA GPUs, ~4.5 GW Google TPU — ns123abc · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27