AMD Instinct MI400 Series / variant
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.
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 |
|---|---|---|---|---|
| 2027 H1 | 5.0k–15k–30k | 0.03M | rumor | Industry reports (mass production Q2 2027) |
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.
Top-of-stack MI400-series die, TSMC N2. 20 PFLOPS FP8 / 40 PFLOPS FP4. 72 GPUs per Helios rack.
AMD Zen 6 host CPU, one per compute tray (18 trays × 1 CPU).
HBM stacks attached to the compute die.
12 stacks per GPU × 72 GPUs = 864 stacks; 31 TB total HBM4, 1.4 PB/s aggregate bandwidth.
Advanced packaging (CoWoS-L) + ABF substrate.
TSMC CoWoS-L 3.5D for each MI455X die + HBM4 stacks.
High-layer-count ABF substrate for OAM module.
Cold plate, CDU, hoses, manifolds.
Distributed cooling layout (independent cold plate + flexible hose) on each GPU and HBM4 stack; extends to DIMMs. Designed for >1200 W heat density per accelerator.
PSUs, busbars, BBU.
Liquid-cooled rack power architecture, ~145 kW per rack (estimated from 72×1.2 kW + overhead).
NICs, switches, cables, CPO modules.
AMD Pensando 800 Gb/s AI NIC, one per MI455X.
AMD Pensando 400 Gb/s DPU for management plane.
UALoE72 scale-up fabric across 72 GPUs in one rack; 3.6 TB/s scale-up bandwidth per GPU.
Rack, sleds, trays, brackets, heatsink fins.
Double-wide open-rack chassis based on Meta 2025 OCP design; ~7000 lb fully populated.
Motherboard, backplane, BMC, signal integrity.
Baseboard management controller per compute tray.
