AMD MI250X

MI250X is AMD’s high-end CDNA 2 GPU.

Specifications

Each MI250X GPU is actually two Graphics Compute Die (GCDs) connected with Infinity Fabric on a single OAM package, so a single GPU appears as two separate GPUs to the operating system and programmer. To be explicit, the following describes a single MI250X package:¹

• 2 GCDs (Graphics Compute Die)
  • 220 CUs (110 per GCD)
    • 14,080 Stream Processors (64 per CU)
    • 880 Matrix Cores (4 per CU)
  • 1.7 GHz (peak)
  • No sparsity
• 128 GB HBM2e (8 stacks)
  • 3.2 TB/s (max)
• 8x AMD Infinity Fabric (D2D)
• 1x16 PCIe Gen4 (H2D)
• 560 W maximum

Performance

The following are theoretical maximum performance in TFLOPS:¹

Data Type	VFMA	Matrix	Sparse
FP64	47.9	95.7
FP32	47.9	95.7
TF32
FP16		383
BF16		383
FP8
INT32
INT8		383
INT4		383

Platforms

The following platforms support MI300X GPUs:

Footnotes

1. AMD CNDA 2 Whitepaper (amd.com) ↩ ↩²