AMD is going strong, also in the professional field. After a year with a lot of competition between RDNA2 and Ampere, now AMD has announced the new Instinct MI100, its new GPU for servers HPC (‘High-performance computing’) manufactured in 7 nanometers that takes advantage of the CDNA architecture. A GPU that promises a 300% improvement over the previous generation, reaching levels never seen before.

The Instinct MI100 bills itself as “the fastest GPU in the world”, reaching 11.5 TFLOPS FP64. To put ourselves in situation, it is the first graphic to overcome the barrier of 10 TFLOPS 64 and clearly surpassing the beastly NVIDIA A100.

Better than NVIDIA A100 40GB by 18%

The most powerful GPU in the world is not for us, but scientists can take advantage of it. It is a graph focused on the world of scientific research. “Aimed directly at the workloads that matter in scientific computing, our latest accelerator, when combined with the open software AMD ROCm platform, is designed to provide scientists and researchers with a superior foundation for their work in HPC.” explains Brad McCredie, VP of Data Center GPU at AMD.

At the gates of the era of exascale, this AMD Instinct MI100 wants to position itself as the clear reference. Its direct rival is the NVIDIA A100, presented last May and also updated today with a new version with up to 80 GB of memory.

But the data offered by AMD seems devastating. This MI100 offers around 18% better performance than the NVIDIA A100 40GB despite having a consumption of 300W, by NVIDIA 400W. A battle that should not change with the new 80 GB A100, which beyond double the memory would offer the same TFLOPS.

The MI100 offers up to 11.5 TFLOPS of peak FP64 performance for HPC and up to 46.1 TFLOPS peak FP32 array performance for AI and machine learning workloads. In FP32, the MI100 achieves 23.1 TFLOPS, down from 19.5 TFLOPS for the Ampere A100.

In previous generations, AMD was well below these numbers, but the move to AMD’s new matrix core technology has resulted in a almost 7-fold increase in theoretical peak performance compared to the previous generation.

Among the companies that work with AMD are Dell, Gigabyte, HP or Lenovo, who will have access to this new Instinct MI100 by the end of the year. According to Bronson Messer, director of Oak Ridge National Laboratory, who was given access to a version of the new chipset, the performance improvement is 2-3 times compared to other GPUs.