Even if Nvidia deserves the most attention today with Hopper and Grace, AMD has also introduced a new computing accelerator. Previously announced without details, the Instinct MI210 PCIe card is now generally available, according to AMD, alongside the larger MI250 and MI250X.
AMD had already presented in November the new generation of the Instinct family with CDNA-2 architecture and the MI210, MI250 and MI250X models, but at the moment it only mentions key data for the two multichip solutions MI250 and MI250X.
The specifications of the Instinct MI210
The Instinct MI210 is the first single chip variant of the new generation and therefore only uses one die. Consequently, with 104 compute units, 6,656 shader units, and three Infinity Fabric links, it’s almost a MI250 cut in half using two of these chips.
Since the maximum clock is also 1,700 MHz, the computing power (FP64 vector: 22.6 TFLOPS) is exactly halved. The same applies to storage: instead of 128 GB at 3.2 TB/s, it’s just 64 GB HBM2e at 1.6 TB/s. The memory is still twice as big as in the old MI100 and also 33 percent faster.
The MI210 is only offered as a PCIe 4.0 plug-in card with a passive cooler and 300 watts of total board power (TBP).
Little message in the shadow of Nvidia Hopper
AMD has today “the general availability of the AMD Instinct MI200 family of accelerators and ROCm 5 software” Announced. Potential buyers are large data centers or supercomputers that focus on scientific calculations, such as in the field of weather and vaccine research or artificial intelligence. Prices are not publicly communicated in this product field .
It is no coincidence that AMD set today to report the earliest availability and unveiling of the MI210, when competitor Nvidia unveiled its new compute accelerator with Hopper GPU and Grace CPU at GTC 2022. The following list serves as a shortcut to more information.
Introvert. Beer guru. Communicator. Travel fanatic. Web advocate. Certified alcohol geek. Tv buff. Subtly charming internet aficionado.