AMD has released the first cDNA-based computing accelerator card Instinct MI100 in 2020, designed for data centers. MI100 has 120 EUs and 7680 stream processors. The graphics card core utilizes 7nm process technology with a core frequency of 1502MHz.
It has 32GB HBM2 video memory, a bit width of 4096-bit, and supports ECC. The computing speed of this computing card exceeds 10TF FP64, and it can reach 184.6 TFLOPs FP16 for artificial intelligence computing. It is one of the fastest computing accelerator cards currently.
The specification s of AMD’s next-generation Instinct MI200 have recently been leaked, code-named ‘Aldebaran’ (Bi Suwu ). The new version of the accelerator card will use the CDNA2 architecture and MCM multi-chip design.
The new version of the computing card upgrades the video memory to HBM2E, and the single-chip video memory capacity can reach 16GB, so the total video memory capacity of the graphics card is expected to reach 64GB.
By looking at the specification, the computing card will support SDMA technology, allowing the system to directly access the video memory. Instinct MI200 is supposed to be launched in the fourth quarter of 2021, and the calculation speed will be twice that of MI100.
(Via)
