NVIDIA Introduces Ampere Architecture-Based Data Center GPU

During GTC2020, the company’s conference held online, NVIDIA CEO Jensen Huang delivered a keynote address to the launch of the NVIDIA A100, a data center GPU with next-generation Ampere architecture.

He said cloud computing and AI trends are driving structural changes in data center design.The NVIDIA A100, which adopts the Ampere microarchitecture, is an end-to-end machine learning accelerator that will boost AI performance by up to 20 times the previous model, Tesla V100, from data analysis to inference. Emphasized.

In addition, the 3rd generation NVLink interconnect implemented with the A100 GPU increases the number of links per GPU and switch, so the 3rd generation NVLink significantly expands the communication bandwidth between GPUs, enables transmission of 50 Gbps for both signals, and error detection and recovery functions. It is said to have improved.

If you look at the specifications of the NVIDIA A100, it uses the TSMC 7nm manufacturing process, the die size is 826mm ² and the number of transistors is 54.2 billion, which is almost twice that of the existing Tesla V100. The GPU memory is 40GB, the memory bandwidth is 1.555TB/sec, and the thermal design power is up to 400W.

The biggest feature of the A100 GPU is that it includes a multi-instance GPU MIG (Multi-Instance GPU) function, that is, the GPU partitioning function. It divides GPU resources and allows parallel processing to other instances. One chip can handle up to 7 instances.

It also announced the Ampere architecture and the A100 GPU, as well as NVIDIA’s parallel computing platform CUDA 11. CUDA 11 accelerates a variety of workloads, including high-performance computation and genetic analysis, 5G communication movie rendering, and deep learning data analysis. Nvidia is also a 3rd-generation integrated AI system with 8 A100s on the GPU, 2 AMD Epic Roma on the CPU, 15TB SSD for storage, 6 NVswitches on the switch chip, and Melanox on the network interface. It also announced the DGX A100, a unit equipped with nine ConnectX-6 200Gbps (ConnectX -6 200Gbps). According to Nvidia, the DGX A100 will show 5 petaflops of computing power in one unit and will cost $199,000.

According to Nvidia, the A100 GPU is already used by various companies and research labs. Microsoft trained the world’s largest language model, Turing NLG, with 17 billion parameters using an Ampere architecture-based GPU. Meanwhile, the location in the video where Jensen Huang gave a keynote speech is the kitchen of his home in California. Related information can be found here .