NVIDIA Tesla V100: Volta GPU, 16GB HBM2 @ 900GB/sec

GTC 2017 - NVIDIA has unveiled its monsterous new Tesla V100 professional graphics card, the first with their next-gen Volta GPU architecture - and NVIDIA's second graphics card with HBM2 technology.

The new Tesla V100 packs 16GB of ridiculously fast HBM2 on a 4096-bit memory bus that provides a huge 900GB/sec of memory bandwidth. Tesla V100 is capable of a huge 15 TFLOPs of single precision (FP32) performance, while it packs 7.5 TFLOPs of double precision (FP64) performance - enough for the largest of datasets in datacenter/AI/deep learning workloads.

Inside, Tesla V100 rocks the company's next-gen Volta GPU architecture with a huge 5120 CUDA cores at 1455MHz boost clock, on the fresh new 12nm manufacturing process - something we reported on a couple of months ago now. If we compare this to the Pascal-based Tesla P100 that was made on the 16nm FinFET prrocess with 16GB of HBM2 at 720GB/sec, and only 3584 CUDA cores in comparison. NVIDIA's new Tesla V100 is an absolute beast.

NVIDIA Tesla V100 Specifications

• GPU: GV100 (Volta)
• CUDA cores: 5120
• Transistors: 21.1 billion
• Node: 12nm
• SMs: 80
• GPU Boost Clock: 1455MHz
• TFLOPs: 15
• VRAM: HBM2
• Memory Bandwidth: 900GB/sec
• Memory Interface: 4096-bit
• TDP: 300W

NVIDIA's new Tesla V100 is a massive GPU with the Volta GPU coming in at a huge 815mm square, compared to the Pascal-based Tesla P100 at 600mm square.

NVIDIA continues to prove it is the king of GPU technology, with CEO and founder Jensen Huang saying during his GTC 17 keynote: "We need to find a path forward for life after Moore's Law", with Huang adding that Volta is "at the limits of photolithography". $3 billion in R&D was spent by NVIDIA, with the company adding it has redesigned Volta's streaming multiprocessor architecture to be 50% more power efficient than Pascal.

The company adds that these tweaks to Volta provide "major boosts in FP32 and FP64 performance in the same power envelope", with NVIDIA adding that the new Tesla V100 also features "tensor cores" that have been built specifically for deep learning, and are capable of 12x the teraflops performance than the Tesla P100.

NVIDIA will launch its new Volta-based Tesla P100 graphics card later this year.