AMDVendor documented
Instinct MI250X
CDNA 2 · CDNA 2 · OAM · 2021
The dual-die CDNA 2 GPU behind Frontier. Exceptional FP64 for HPC; usable for training on ROCm-supported frameworks.
HPCLLM trainingLLM inference
Precision fingerprint
6432t3216BF168i8
Memory
128 GB
HBM2e
Bandwidth
3.2 TB/s
peak
TDP
560 W
liquid
Max model
~50B
FP16, planning est.
Compute throughput
| FP64 | 95.7 TFLOPS |
| FP32 | 95.7 TFLOPS |
| TF32 | — |
| FP16 | 383 TFLOPS |
| BF16 | 383 TFLOPS |
| FP8 | — |
| INT8 | 383 TFLOPS |
Platform & software
InterconnectInfinity Fabric — 800 GB/s
PCIePCIe 4.0 x16
Coolingliquid
MIGNot supported
PartitioningDual-GCD (2 logical GPUs)
VirtualizationSR-IOV
FrameworksROCm, RCCL, MIOpen
Availabilitybare-metal
Known limitations
- ·No FP8
- ·Dual-GCD design appears as two GPUs to schedulers