At today’s Next Horizon event, AMD announced their enterprise class MI50 and MI60 gpus. These gpus are optimised for datacenter applications and support hardware virtualisation. Not only are they the first gpus produced on TSMC’s new 7 nm process, they are also the first gpus to incorporate the new pci-e 4.0 standard.
Built for high performance compute workloads such as machine learning using the open source ROCm software platform.
Both cards support AMD’s Infinity Fabric Link technology, which can be used to directly connect up to 2 GPU clusters of 4 GPUs in a single server at up to 200 GB/s.
MI50 and MI60 specifications
| Radeon Instinct MI60 | Radeon Instinct MI50 | |
|---|---|---|
| Memory | 32 GB HBM2 | 16 GB HBM2 |
| Bandwidth | 1 TB/s | 1 TB/s |
| Architecture | Vega20 | Vega20 |
| Compute Units | 64 | 60 |
| Streaming Processors | 4096 | 3840 |
| Max. Clock Speed | 1800 MHz | 1746 MHz |
| Half Precision (FP16) | 29.5 TFLOPs | 24.6 TFLOPs |
| Single Precision (FP32) | 14.7 TFLOPs | 13.4 TFLOPs |
| Double Precision (FP64) | 7.4 TFLOPs | 6.7 TFLOPs |
| Inference (INT8) | 59 TOPs | 53.6 TOPs |
| Power Usage (TDP) | 300 W | 300 W |
| Cooling | Passive (Datacenter) | Passive (Datacenter) |