The Open-Source AI Stack
RSS
All hardware

Hardware · Datacenter

AMD Instinct MI325X

AMD

Memory refresh of the MI300X to 256 GB HBM3e at higher bandwidth; same CDNA3 dense compute.

×8 over Infinity Fabric Compute units 1307 TF FP16 6.00 TB/s memory bus 256 GB HBM3e (VRAM)
Datacenter. Bus width tracks bandwidth (6.00 TB/s, sets decode speed); the box tracks capacity (256 GB, sets what fits). Drawn per unit; a typical node is 8×.

Specs

Memory
256 GB HBM3e
Bandwidth
6.00 TB/s
FP16 dense
1307 TFLOPS
FP8 dense
2615 TFLOPS
Power
1000 W
Form factor
oam
Interconnect
infinity-fabric
Released
2024-10

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

Model Fits? Decode ceiling
Llama 3.1 8B Instruct yes ~8105 tok/s
Llama 3.3 70B Instruct yes ~1005 tok/s
Qwen 2.5 72B Instruct yes ~976 tok/s
DeepSeek-V3 yes ~1942 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