AMD · 2026-2027

AMD Instinct MI400 Series

AMD's next-generation AI accelerator family unveiled at CES 2026: TSMC N2 logic with potential partial shift to Samsung Foundry SF2P (decision early 2026), CoWoS-L 3.5D packaging, 432 GB HBM4 per GPU, and rack-scale liquid cooling via the Helios open-rack architecture (built on Meta 2025 OCP design). Available across three accelerator SKUs (MI430X for sovereign/HPC, MI440X for server training+inference, MI455X for rack-scale). Helios rack: 72× MI455X + 18× EPYC Venice CPUs delivering 2.9 EFLOPS FP4.

主要 variants

點任一張卡片進入該 variant 的完整 BOM 與供應商分析。

2 個 variant並排比較 BOM →

AMD Instinct MI450

2026 H2

Server-scale MI450 accelerator (UBB8 8-GPU baseboard). 432 GB HBM4 per GPU, 19.6 TB/s HBM bandwidth, 20 PFLOPS FP8 / 40 PFLOPS FP4. Oracle named as 50,000-GPU launch customer for 2026 (Tom's Hardware).

Liquid-cooled1 kWTSMC N2

2026 H2rumor

~50k 台 (30k–70k) · ASP ~0.0M USD

查看 BOM →

AMD Instinct MI455X (Helios Rack)

2027 H1

Rack-scale Helios UALoE72 system: 72× MI455X + 18× EPYC Venice CPUs across 18 compute trays in a double-wide open rack (~7000 lb, Meta OCP 2025 design). 31 TB HBM4 total, 1.4 PB/s aggregate bandwidth, 2.9 EFLOPS FP4 / 1.4 EFLOPS FP8. Networking via AMD Pensando Vulcano 800 NIC + Salina 400 DPU. Engineering samples H2 2026; mass production Q2 2027.

Liquid-cooled145 kWTSMC N2

2027 H1rumor

~15k 台 (5.0k–30k) · ASP ~0.0M USD

查看 BOM →