Announced at GTC 2017, the Tesla V100 is an enterprise-level processor powered by the Volta GV100 GPU: the first chip in the world built with a 12nm FFN process.
A single Volta GV100 packs in 21 billion transistors, 5120 CUDA cores, 320 texture units and a 4096-bit HBM2 memory interface with a boost clock speed of 1455MHz. It's equipped with 640 Tensor Cores capable of providing 120 teraflops of tensor operations. (And yes, it will totally play Crysis - one day.)
Boasting 16GB HBM2 operating at 900 GB/s, the Volta architecture is purportedly four times faster than NVIDIA's top-of-the-range Pascal GPUs in peak teraflops. According to NVIDIA CEO Jensen Huang, thousands of NVIDIA engineers spent over three years crafting Volta for a total development cost of $3 billion.
In addition to being NVIDIA's most powerful GPU it is also the biggest: the die size is 815mm2 which is roughly the same size as an Apple Watch. (By contrast, the Tesla P100 measures 600mm.)
As mentioned, the chipset will debut in the Tesla V100 range, which are data center GPUs built especially for deep learning and artificial intelligence. Here's a look at the specs:
NVIDIA is pitching the Tesla V100 at all industries that are keen to get involved in AI. (i.e. - it's not just for niche researchers and data scientists.) However, they won't come cheap: a Volta-based DGX system with eight Tesla V100s inside will set you back a colossal $149,000.
"Artificial intelligence is driving the greatest technology advances in human history," Huang proclaimed during his GTC keynote. "It will automate intelligence and spur a wave of social progress unmatched since the industrial revolution." Bold words.
So is any of this relevant to gamers and PC enthusiasts? Almost certainly. While the Tesla V100 is strictly aimed at data centres, Volta-based consumer graphics cards will probably be appearing in the future.
During the GTC keynote, Huang showed off a demo of Kingsglaive: Final Fantasy XV rendered with Volta processors. It contained some of the most photorealistic fabrics I've ever seen. NVIDIA's commitment to deep learning GPUs can only mean good things for GeForce graphics cards. Much like the Pascal-based Titan X, we will hopefully see Volta trickle down in one form or another in the future.
Lifehacker attended GTC 2017 as a guest of Nvidia.