AMD Instinct MI400 Series / variant
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).
區間 + 信心水準 + 來源；不同分析師、不同時點的預估會並列。
| 期間 | 出貨量 (low–mid–high) | ASP (USD) | 信心 | 來源 |
|---|---|---|---|---|
| 2026 H2 | 30k–50k–70k | 0.02M | rumor | Tom's Hardware (Oracle 50k deployment 2026) |
每行 = 單一機台的零組件 slot。share % 是該供應商在此 component 的估計份額。
GPU / CPU dies — the headline silicon.
CDNA 5 architecture, TSMC N2 logic, 20 PFLOPS FP8 / 40 PFLOPS FP4 per die.
AMD Zen 6 server CPU (host), TSMC N2 class.
HBM stacks attached to the compute die.
12-stack 432 GB total per GPU, 19.6 TB/s aggregate bandwidth.
Advanced packaging (CoWoS-L) + ABF substrate.
TSMC CoWoS-L 3.5D advanced packaging for GPU + HBM4 stacks.
High-layer-count ABF substrate for OAM module.
Cold plate, CDU, hoses, manifolds.
Independent cold plate + flexible hose per GPU and HBM4 stack; thermal density >1200 W per accelerator.
NICs, switches, cables, CPO modules.
AMD in-house 800 Gb/s AI NIC (Pensando division).
AMD in-house 400 Gb/s DPU for management plane.
Motherboard, backplane, BMC, signal integrity.
Baseboard management controller for AI server tray.
