AMD today announced its first GPU based on computer-oriented CDNA architecture.
AMD announces INSTINCT MI100 – “World’s Fastest HPC GPU”
The INSTINCT MI100 drops the Radeon naming. AMD made a similar decision to NVIDIA who dropped Quadro and Tesla for its HPC accelerators and workstation cards. The MI100 features a new GPU featuring 120 Compute Units (7680 Stream Processors), which is 40 CUs more than Navi 21 (Big Navi).
The MI100 is equipped with 32GB of HBM2 memory (four HBM stacks) with a maximum theoretical bandwidth of 1.23 TB/s. The product uses PCI Express Gen4, similar to the AMD Navi series and Vega 20.
AMD claims that the CDNA-based GPU will have twice as high compute density as the last generation. It will feature an all-new matrix core technology for HPC and AI workloads. AMD says that it will be nearly 7X on AI workloads with Mixed Precision and FP16, up to 4X faster GPU-to-GPU performance with Infinity Architecture vs PCIe 4.0.
The AMD Instinct MI100 will have peak FP64 (double-precision) performance of 11.5 TFLOPs and FP32 (single-precision) of 23.1 TFLOPs. With the MATRIX technology, the FP32 figure will go up to 46.1 TFLOPs and FP16 (half-precision) compute performance should reach 184.6 TFLOPs. AMD also lists the BFLOAT16 performance of 92.3 TFLOPs.
Introducing AMD Instinct™ MI100 accelerator: First data center GPU to surpass the 10TF FP64 barrier.
AMD Instinct™ MI100 accelerator, the world’s fastest HPC GPU, delivers up to 11.5 Teraflops peak double precision (FP64) performance. Engineered from the gro…