The Open-Source AI Stack
RSS
All hardware

Hardware · Datacenter

AMD Instinct MI355X

AMD

CDNA4 generation; shipping. Adds an FP4 path. AMD prints dense and sparse side by side; dense values used here.

×8 over Infinity Fabric Compute units 2517 TF FP16 8.00 TB/s memory bus 288 GB HBM3e (VRAM)
Datacenter. Bus width tracks bandwidth (8.00 TB/s, sets decode speed); the box tracks capacity (288 GB, sets what fits). Drawn per unit; a typical node is 8×.

Specs

Memory
288 GB HBM3e
Bandwidth
8.00 TB/s
FP16 dense
2517 TFLOPS
FP8 dense
5033 TFLOPS
Power
1400 W
Form factor
oam
Interconnect
infinity-fabric
Released
2025-06

What it runs (8× unit, Q4_K_M, 4K context)

Model Fits? Decode ceiling
Llama 3.1 8B Instruct yes ~10807 tok/s
Llama 3.3 70B Instruct yes ~1340 tok/s
Qwen 2.5 72B Instruct yes ~1302 tok/s
DeepSeek-V3 yes ~2590 tok/s

Ceiling is the theoretical rooflineruntimeA performance model that bounds throughput by either compute or memory bandwidth, whichever is the limiting resource for an operation's arithmetic intensity. Open full entry ; open the explorer to set quant, context, and runtime and see the realistic range.

Sources