AMDVendor documented
Instinct MI325X
CDNA 3 · CDNA 3 (refresh) · OAM · 2024
MI300X refreshed with 256 GB HBM3e and 6 TB/s — the memory-capacity leader, aimed at fitting the largest models per GPU.
LLM trainingLLM inferenceFine-tuningHPCMultimodal
Precision fingerprint
6432t3216BF168i8
Memory
256 GB
HBM3e
Bandwidth
6 TB/s
peak
TDP
1000 W
air or liquid
Max model
~110B
FP16, planning est.
Compute throughput
| FP64 | 163.4 TFLOPS |
| FP32 | 163.4 TFLOPS |
| TF32 | — |
| FP16 | 1.31 PFLOPS |
| BF16 | 1.31 PFLOPS |
| FP8 | 2.62 PFLOPS |
| INT8 | 2.62 PFLOPS |
Platform & software
InterconnectInfinity Fabric — 896 GB/s
PCIePCIe 5.0 x16
Coolingair or liquid
MIGNot supported
PartitioningCompute + memory partitioning, SR-IOV
VirtualizationSR-IOV
FrameworksROCm, RCCL, vLLM (ROCm), MIOpen
Availabilitybare-metal
Known limitations
- ·1000 W raises cooling requirements
- ·ROCm ecosystem still maturing