“Deep learning has insatiable demand for processing power,” said NVIDIA’s CEO Jensen Huang. “Thousands of NVIDIA engineers spent more than three years crafting Volta to help meet this need, enabling the industry to realise AI’s life changing potential.”
Volta is said to offer five times the peak TFLOPS performance of Pascal, the current NVIDIA GPU architecture, and 15 times the performance of the two year old Maxwell architecture.
The architecture features 640 Tensor Cores, with CUDA cores and Tensor Cores paired to provide what the company calls the performance of an AI supercomputer in a single GPU. Also included is the NVLink high-speed interconnect, linking GPUs and CPUs, with up to twice the throughput of the prior generation NVLink. Memory operations are supported by the 900Gbyte/s HBM2 DRAM interface, developed in collaboration with Samsung.
The company also announced the Volta based Tesla V100 data centre GPU, targeted at AI inferencing and training, as well as for accelerating high performance computing and graphics workloads.
