A "kitchen keynote" by Nvidia CEO Jensen Huang reveals the first product based on next-generation Ampere architecture-- the A100, a datacentre GPU built for data analytics, scientific computing and cloud graphics.
The company claims the A100 is the largest 7nm-based processor, carrying over 54 billion transistors. It includes 3rd generation Tensor Cores with TF32, a math format to accelerate single-precision AI training, and uses structural sparsity acceleration for higher performance. A multi-instance GPU (MIG) allows customers to partition a single A100 into as many as seven independent GPUs, each with own resources, while 3rd generation NVLink technology doubles the high-speed connectivity between GPUs, turning A100 servers into one giant GPU.