On the evening of May 14th, at the GTC 2020 online launch, NVIDIA simply released the old Huang’s speech video in a matter of minutes, officially launched the Ampere architecture GPU, 7nm process, 54 billion transistors, 20 x AI computing power, 5 major technological innovations, in short, the new nuclear bomb came.
NVIDIA has not officially released the details of the ampgpu architecture, but like the turing GPU, Huang said it was the biggest performance leap in nVIDIA’s eight-generation GPU history.
The first Ampere-based GPU is the Tesla A100 Acceleration Card, which, according to NVIDIA, brings five major technological innovations:
1, the new ampere GPU architecture, 54 billion transistors, the world’s largest 7nm processor.
2, the third generation of Tensor Core AI core, support TF32 operations, without any code changes can improve AI performance 20 times, but also support FP64 double-precision operations, compared with HPC applications bring 2.5 times performance.
3, Multi-instance GPU Multi-Instance GPU: MIG, an innovative technology that divides a GPU into seven separate GPUs, providing different operations for different targets and maximizing computational efficiency.
4, NVLink 3.0: The performance of the next-generation GPU bus doubles to provide more efficient performance scaling in server applications.
5, structural sparseness: This new technology uses the inherent sparseness of AI operations to double performance.
These five technological innovations make the Tesla A100 accelerated card ideal for demanding workloads, not only for AI reasoning, AI training, but also for scientific simulation, AI conversation, genomics, high-performance data analysis, seismic modeling, and financial computing.
At the same time, NVIDIA also announced the Tesla A100-based DGX A100 supercomputing, with eight Tesla A100 acceleration cards with performance up to 5PFLOPS, Alibaba Cloud, AWS Cloud, Google Cloud, Microsoft Azure, Oracle and Tencent Cloud will all launch DGX A100-based cloud services.
The DGX A100 is now on the market immediately after its launch, unlike the Tesla V100, which has been used as an overcomputing solution in several laboratories and supercomputing centers in the United States and Germany.