NVIDIA launches Tesla V100 GPU based on Volta architecture, with 16GB HBM2 memory and 900GB/sec bandwidth.

NVIDIA has unveiled its new Tesla V100 professional graphics card at GTC 2017, the first featuring the company's next-generation Volta GPU architecture and NVIDIA's second graphics card with HBM2 technology. The Tesla V100 features 16GB of HBM2 memory on a 4096-bit memory bus, providing 900GB/sec of memory bandwidth. The card delivers 15 TFLOPs of single precision (FP32) performance and 7.5 TFLOPs of double precision (FP64) performance, making it suitable for the largest datasets in datacenter, AI, and deep learning workloads.

The Tesla V100 is built on NVIDIA's next-generation Volta GPU architecture, featuring 5120 CUDA cores running at a 1455MHz boost clock on a 12nm manufacturing process. Compared to the Pascal-based Tesla P100, which was manufactured on a 16nm FinFET process with 16GB of HBM2 at 720GB/sec and only 3584 CUDA cores, the Tesla V100 represents a significant advancement. The Volta-based card measures 815mm square, compared to the Tesla P100's 600mm square.

During his GTC 17 keynote, NVIDIA CEO and founder Jensen Huang stated, "We need to find a path forward for life after Moore's Law," and added that Volta is "at the limits of photolithography." NVIDIA invested $3 billion in R&D for the development and has redesigned Volta's streaming multiprocessor architecture to be 50% more power efficient than Pascal. These improvements provide "major boosts in FP32 and FP64 performance in the same power envelope," according to the company.

The Tesla V100 features "tensor cores" that have been built specifically for deep learning applications and are capable of delivering 12x the teraflops performance compared to the Tesla P100. NVIDIA will launch its new Volta-based Tesla P100 graphics card later this year.