AMD MI250 (X): Aldebaran as MCM with up to 47.9 teraflop / s and 560 watt TDP

AMD has introduced the Instinct MI250 (X). The two HPC accelerators with the code name Aldebaran have a high computing power on paper, but also have to be cooled in a complex manner.

Nov 9, 2021 - 08:27
 10
AMD MI250 (X): Aldebaran as MCM with up to 47.9 teraflop / s and 560 watt TDP

AMD presented news about its server business and part of it was the MI200 series based on the CDNA2 architecture, where the main focus is on the Instinct MI250X. For them, AMD promises a 4.9 times better performance than competing products under FP64 loads. But the top model, the Instinct MI250X, should also convince in other scenarios such as half-precision in AI tasks. AMD offers 3 cards of the type "Aldebaran", the Instinct MI210, the Instinct MI250 and said Instinct MI250X. The technical data is still missing for the smallest card. 

Technically, the Instinct MI250 (X) relies on two compute chips (MCM, Multi Chip Module) each with a clock rate of 1.7 GHz and eight stacks of HBM2E memory with a total of 128 GiByte capacity at 1.6 GHz. AMD names 47.9 teraflop / s computing power with 32 and 64 bit and 308 teraflop / s with 16 bit and 3.2 TB / s memory transfer capacity. The main difference between the Mi250 and MI250X are the execution units with 13,312 FP32 units (202 CU) for the MI250 and 14,080 units (220 CU) for the MI250X, which in the end amounts to around 2.5 teraflop / s computing power. In addition, the two models differ in the fact that the non-X model has six instead of eight infinity links and offers no memory coherence. 

Models Compute Units Stream Processors FP64 | FP32 Vector (Peak) FP64 | FP32 Matrix (Peak) FP16 | bf16(Peak) INT4 | INT8(Peak) HBM2eECCMemory Memory Bandwidth Form Factor
AMD Instinct™ MI250x 220 14,080 Up to 47.9 TF Up to 95.7 TF Up to 383.0 TF Up to 383.0 TOPS 128GB 3.2 TB/sec OCP Accelerator Module
AMD Instinct™ MI250 208 13,312 Up to 45.3 TF Up to 90.5 TF Up to 362.1 TF Up to 362.1 TOPS 128GB 3.2 TB/sec OCP Accelerator Module

The high raw output of the chips manufactured in 6 nm with 58 billion transistors has its price in the form of a TDP of 560 watts (liquid-cooled) or 500 watts (air-cooled). The cards are offered in the form factor OAM (OCP Accelerator Module). The small model, MI210, is sold as PCIE. The direct competitor is the A100 from Nvidia, which is marketed as SXM and offers significantly less performance on paper, but is the smaller chip (7,680 FP32 shader) with a TDP of 400 watts. Nvidia names 46.1 teraflop / s computing power for 32 and 64 bit and 184.6 / 92.3 teraflop / s with FP16 / BF16. Independent tests must show what is more efficient in the end and will also depend on the task. Depending on the calculation, AMD calculates double or quadruple performance with 87 percent more power consumption. AMD Instinct MI250 (X) is available immediately through partner HP. The MI210 plug-in card will appear shortly. 

By: Amber V