AMD Instinct MI400 Series GPU Architecture(MI455X)
Speakers: Alan Smith, Maiyuran Subramaniam(AMD)
Venue: Hot Chips 2026 Conference
PDF: raw/papers/HC2026_AMD_Instinct_MI455X.pdf(非公开本地路径)
芯片拆解。系统见同日 UALoE。对照 Rubin NVLink 6。封装落到 Hybrid Bonding 与 DRAM。
Helios 一页摘要
72 GPUs/rack;2.9 Exaflops AI compute;31 TB HBM4;1.7 PB/s HBM4 bandwidth;260 TB/s scale-up;43 TB/s scale-out。Compute tray:4 Instinct MI455X EAM;单路 EPYC 9006 SP7(Venice);36 UALoE links (x2) per EAM;UALoE 1.8 TB/s/dir per GPU;Infinity Fabric to CPU 128 GB/s/dir per GPU;每 EAM 最多 3× Pensando Vulcano 800 AI NIC,经 UALink128 (x8)、128 GB/s/dir per NIC。
MI455X chiplets
- 2× FCD, N3P — 192-channel HBM4;192 MB Global L2
- 2× IOD, N3P — 2× PCIe Gen6 或 3× AI-NICs via UALink;Infinity Fabric 256 GB/s bi-dir;72 UALoE lanes, 3.6 TB/s bi-dir
- 8× XCD, N2 — 256 total active Work Group Processors
- 12× HBM4 — 432 GB at 23.3 TB/s
封装:3D hybrid-bonded XCDs;CoWoS-L;SOIC + HBM4。概念 SoC 框图页几乎无字(未知)。slide 上 192 MB Global L2 与 vs-MI355X 的 96 MB L2 都在片上;暂按 2×96 MB FCD 读。
Cache vs MI355X:VGPR 128 KB 不变,标 2.0× VGPR per SIMD;scalar 8 KB vs 3.2 KB;LDS 384 KB vs 160 KB(2.0× LDS per WGP/CU);L2 96 MB vs 4 MB L2 + 256 MB Infinity Cache(1.5× per L2 vs Cache);HBM 432 GB HBM4 vs 288 GB HBM3E(2.9× total HBM)。Broadcast arbitrator:「up to 4× BW amplification」。
算力与效率
Peak vs MI355X(AMD Performance Labs, June 2026,endnote MI400-006):OCP MXFP4 40.26 PF(up to 4X);OCP MXFP6 20.13 PF(up to 2X);OCP MXFP8/FP8 20.13 PF(up to 4X);Matrix FP16/BF16 5.03 PF(up to 2X);Vector FP16 315 TF(up to 2X);Matrix/Vector FP32 315 TF(up to 2X)。MXFP8/6/4,block-scale 16/32;native Wave32。
效率:Tensor Data Mover(async Global↔LDS);work-group cluster + L2 multicast + L2 prefetch;split/named barriers。每 GPU 专用 DMA:topology-aware,流量自动亲和 UALoE,「software … oblivious to data placement」;「scales across 72 GPUs」。
实测 vs MI355X(ROCm.ai,July 2026):MLA decode FP8 20 TB/s(3.8×,单卡);FP4 compute 20 PF(3.3×,单卡 AITER GEMM);scale-up BW 3.2 TB/s(3.5×,4× MI455X vs 8× MI355X,16 GB transfer);scale-out BW 190 GB/s(2×,Vulcano 800 vs Pollara 400)。能耗:estimated 2.4× AI energy efficiency vs MI355X;2030 目标 rack-scale efficiency 20×。单卡 TDP / 时钟:未知。
Citations
[1] raw/papers/HC2026_AMD_Instinct_MI455X.pdf(非公开本地路径) — Smith / Subramaniam, Hot Chips 2026
[2] hc2026-amd-instinct-mi455x.md — 结构化摘录