Not only EPYC Milan-X, AMD introduced a third accelerator from the Instinct family called MI200 Instinct MI210 which is added to the MI250 and MI250X solutions presented last November.
The company had anticipated it, so not an unexpected novelty: the first solutions are in fact available in the OCP Accelerator Module (OAM) format, while Instinct MI210 in the classic “card” format, with PCI Express connectionintended for more generic and traditional servers.
Not the only difference between Instinct MI210 and the other two solutions: the newcomer the precise half of Instinct MI250in fact composed of a single GPU 6-nanometer-based architecture CDNA 2 with 104 Compute Unit attive per 6656 stream processor and FP64 vector computing power of up to 22.6 teraflops. The card has 64 GB of HBM2E memory with a bandwidth of 1.6 TB / s. The TDP of 300Wabout 260W lower than that of the models already presented.
The MI250 and MI250X solutions are the first based on two dies (AMD talks about GCD, Graphic Compute Die), each with a maximum of 110 CU: MI250X has 220 CU for a total of 14080 stream processors. The MI250, on the other hand, sees some units disabled (104 CUs active) and stops at 208 CUs and 13312 total stream processors.
| Compute Unit | Stream processor | FP64|FP32 Vector (picco) | FP64|FP32 Matrix (picco) | FP16 | bf16 (peak) | INT8 (peak) | HBM2E ETC. |
||
| AMD Instinct MI210 | 104 | 6656 | Fino a 22,6 TF | Fino a 181 TF | Fino a 181 TF | Up to 181 TOPS | 64GB | Up to 1.6 TB / s |
| AMD Instinct MI250 | 208 | 13312 | Fino a 45,3 TF | Fino a 362,1 TF | Fino a 362,1 TF | Up to 362.1 TOPS | 128GB | 3.2TB/s |
| AMD Instinct MI250X | 220 | 14080 | Fino a 47,9 TF | Fino a 383 TF | Fino a 383 TF | Up to 383 TOPS | 128GB | 3.2TB/s |
AMD competes the new Instinct MI210 with NVIDIA’s A100 40GB, highlighting its far superior performance in a variety of workloads. Instinct MI210, like the proposals that preceded it, completely compatible with the open source platform ROCm 5.0.