AMD Instinct MI250X
Datacenter
The MI250X features dual Graphics Compute Dies (GCDs) with 128GB HBM2e memory on CDNA 2 architecture. Powers the Frontier supercomputer and excels at HPC and AI training.
VRAM
128 GB
Memory
HBM2e
Bandwidth
3277 GB/s
TDP
560W
Suitable Workloads
Large Language Models
Training and inference for models like GPT-4, Llama 70B+
High-Throughput Inference
Optimized for batched inference workloads
Enterprise Deployment
Designed for 24/7 datacenter operations
Key Highlights
- Built on CDNA 2 architecture
- 880 AI accelerator cores for tensor operations
- 14,080 compute cores for parallel processing
- OAM form factor
Memory
VRAM128 GB
Memory TypeHBM2e
Memory Bandwidth3277 GB/s
Compute Performance
FP32 (Single Precision)47.9 TFLOPS
FP16 (Half Precision)383 TFLOPS
BF16383 TFLOPS
INT8383 TOPS
Architecture
ArchitectureCDNA 2
Compute Cores14,080
AI Accelerators880
Release Year2021
Power & Physical
TDP560 W
Max Power560 W
Form FactorOAM
PCIe GenerationGen4
Notes
Dual-GCD design for HPC workloads. Powers Frontier supercomputer.
Models That May Fit on AMD Instinct MI250X
Estimates based on INT8 quantization. Actual fit depends on framework and batch size.
Explore More
Added Jan 27, 2026
Last updated: Jan 27, 2026
Start building with the right model.
From model selection to production, one platform, no fragmentation.