AMD Instinct MI300X Platform
Overview
Performance Metrics
| Precision | Peak (TFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 654 | 82 | — |
| FP32 | 1,307 | 163 | — |
| TF32 | 5,230 | 654 | — |
| FP16 | 10,459 | 1,307 | — |
| BF16 | 10,459 | 1,307 | — |
| FP8 | 20,919 | 2,615 | — |
| INT8 | 20,919 | 2,615 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- Instinct MI300X 192GB
- Count
- 8
- Memory per GPU
- 192 GB
- Bandwidth per GPU
- 5.3 TB/s
Power & Efficiency
- Aggregate Power
- 6.0 kW
- TDP per GPU
- 750 W
- FP8 Efficiency
- 3.49 TFLOPS/W
- Aggregate Bandwidth
- 42.4 TB/s
System Details
Interconnect
4th Gen AMD Infinity Fabric, eight fully connected OAM modules on an OCP UBB 2.0 baseboard, 7x 128 GB/s links per GPU, 896 GB/s ring-of-8 aggregate
Notes
AMD's eight-GPU building block for CDNA 3, integrating eight MI300X OAM modules on an industry-standard OCP universal baseboard (UBB 2.0), each linked to the other seven over 128 GB/s Infinity Fabric links for 896 GB/s of aggregate ring bandwidth, plus one PCIe Gen 5 x16 host link per GPU. AMD's platform datasheet gives 1.5 TB of HBM3 (stored as 8 x 192 GB = 1.536 TB) and 5.3 TB/s per GPU, so 42.4 TB/s across the board, with a 750 W maximum board power per GPU (6.0 kW for the eight GPUs, not counting the host). Compute: 2,432 compute units, 155,648 stream processors at up to 2,100 MHz. AMD sells it through solution partners rather than directly. Precision figures are eight times the per-GPU MI300X row, which AMD's one-decimal platform figures round, except FP32 and FP64, where AMD's platform figures are printed more precisely and are stored as printed.