According to a tweet from an engineer working on the LUMI supercomputer, AMD has quietly begun shipments of its Instinct MI210 PCIe card for high-performance computing (HPC). The accelerator can be used inside both servers and workstations, and will compete against Nvidia's A100 PCIe as AMD's new flagship accelerator card. Perhaps the most interesting thing, however, is that although AMD yet has to formally confirm the actual specifications of the product, the researcher shared a few of the unreleased specs a bit early.
Playing with one @AMDInstinct MI210, BabelStream with HIP is around 40% more than MI100, seems great. #HPC #GPUDecember 3, 2021
From the tweet, we can surmise that AMD has already started shipments of the Instinct MI210 PCIe cards to interested parties, according to George Markomanolis, an engineer from the CSC – IT Center for Science in Finland, where the 0.55 ExaFLOPS LUMI supercomputer is being built. The MI210 has been announced at a live event, but AMD didn't release the official specifications.
Yes, 104 CUs, 64 GB HBM2eDecember 3, 2021
As it turns out, the Instinct MI210 has little to do with the dual-GCD Instinct MI250X. The compute GPU card has 104 compute units, which translates into 6656 stream processors, and 64GB of HBM2e memory. Given the specifications of the product, it looks like it uses only one GCD and therefore is not meant to achieve extreme performance. Meanwhile, with 64GB of memory and CDNA 2 architecture, the MI210 can indeed offer a formidable combination of performance and capabilities compared to AMD’s own Instinct MI100 or Nvidia’s A100 PCIe cards.
By contrast, the flagship Instinct MI250X accelerator with 128GB of HBM2e memory features 14,080 stream processors and 47.9 FP64 TFLOPS performance, comes in an open accelerator module (OAM) form-factor, and consumes up to 550W, according to rumors.
At present, it is impossible to tell how fast AMD's Instinct MI210 is. Meanwhile, it will certainly be interesting to see how AMD's MI210 stacks up against Nvidia's A100 PCIe GPU.