AMD has released the Instinct MI100 Computing Accelerator. For the high-performance (HPC) graphics category, the company has decided to ditch the Radeon brand.
The AMD Instinct MI100 accelerator is equipped with a new GPU that includes 120 compute units (7,680 stream processors). The GPU contains 40 more compute units than is available in the Navi 21 (Big Navi) GPU. The device also includes 32GB of HBM2 memory with a maximum theoretical bandwidth of 1.23TB / s. The PCI Express Gen4 interface is used to connect to the system.
AMD claims that GPUs based on the new CDNA architecture will have twice the previous generation’s compute density. The chip includes completely new core matrix technology for HPC and AI workloads. AMD says this solution will deliver nearly 7x the performance boost for mixed-precision AI and FP16 workloads.
The AMD Instinct MI100 accelerator has a peak double precision (FP64) performance of 11.5 teraflops and a single-precision (FP32) performance of 23.1 teraflops. With MATRIX technology, FP32 performance can be increased to 46.1 teraflops, and half-precision (FP16) performance can reach 184.6 teraflops. Floating-point performance (BFLOAT16) is reported at 92.3 teraflops.