AMD Instinct MI355X Platform
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.6 | 78.6 | — |
| FP32 | 1.3 | 157.3 | — |
| FP16 | 20.1 | 2516.6 | — |
| BF16 | 20.1 | 2516.6 | — |
| FP8 | 40.3 | 5033.2 | — |
| FP4 | 80.5 | 10066.3 | — |
| INT8 | 40.3 | 5033.2 | — |
| FP6 | 80.5 | 10066.3 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- Instinct MI355X OAM
- Count
- 8
- Memory per GPU
- 288 GB
- Bandwidth per GPU
- 8.0 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- 1400 W
- FP8 Efficiency
- —
- Aggregate Bandwidth
- 64.0 TB/s
System Details
Interconnect
4th Gen AMD Infinity Fabric, eight fully connected OAM modules on an OCP UBB design
Notes
AMD's eight-GPU building block, integrating eight fully connected MI355X OAM modules on an industry-standard OCP universal baseboard over fourth-generation Infinity Fabric. AMD publishes the platform totals of 2.3 TB of HBM3E and 64 TB/s of aggregate memory bandwidth, both of which are exactly eight times the per-GPU figures, which is why the precision values here are also eight times AMD's published per-GPU specification rather than a separate platform table, since AMD provides none. Sparse figures appear only for the four datatypes where AMD publishes a structured-sparsity number, which are FP8, FP16, BF16 and INT8; AMD gives no sparsity figure for MXFP4 or MXFP6, so neither does this row. Supported in both air-cooled UBB servers and direct liquid cooled platforms. AMD publishes no platform power figure.