AMD introduces the Instinct MI100 computing accelerator for supercomputers
AMD has released one of the most productive computing accelerators for computers that are used in the field of scientific research. According to the company, the new product will provide greater attractiveness in terms of price and performance compared to the NVIDIA solution.
The new Instinct MI100 received 120 execution units (Compute Units). The declared performance of the accelerator is 10 TFlops for calculations with double precision (FP64). In the FP32 discipline, which is typical for operations involving artificial intelligence and neural networks, the Instinct MI100 demonstrates performance of 46.1 TFlops.
The graphics accelerator is equipped with 32 GB of HBM2 memory running at 1.2 GHz with a bandwidth of 1.23 TB/s. Support for the current PCI-Express Gen 4.0 interface is also announced, while the new product is backward compatible with the previous generation bus. It also supports combining up to four GPUs in a cluster via the AMD Infinity Fabric interface — in this case, the total throughput will be up to 276 GB/s.
The commercial release of the AMD Instinct MI100 is expected before the end of 2020. Among the potential customers of the brand's new products are Dell, Gigabyte, HP, and SuperMicro.