Nvidia’s New GV100 Brings Volta to the Quadro Family

Nvidia has kept its Volta architecture confined to the high-performance computing (HPC) space for years now, with only a handful of exceptions, like the $3,000 Titan V GPU. Now, the company is extending Volta into new markets, this time by bringing the card to workstations.

The new GV100 packs 5,120 CUDA cores, 640 tensor cores, 320 texture units, and 128 ROPS (5120:320:128 in standard GPU parlance). It’s also packing twice the HBM2 (32GB, compared with 16GB for the GP100), and its RAM clock is higher, at 1.7GHz compared with 1.4GHz. The GPU is manufactured on TSMC’s 12nm FFN process (previous Quadro cards were built on the 16nm node). Keep in mind, 12nm isn’t a new node for TSMC despite the nomenclature — both GlobalFoundries and TSMC are using 12nm to reflect an optimized 14/16nm, not a brand-new deployment. Nvidia thinks the tensor cores could be useful in some professional visualization tests, though we’re not sure if any applications currently make use of the feature.

The Quadro GV100 will offer a pair of NVLink connectors for quick low-latency connections and are the major GPUs that’ll be used for Nvidia’s RTX ray tracing technology. It’ll be interesting to see how this evolves over time. I’ve not had a chance to dig into RTX much, but ray tracing is a topic we’ve covered in the past at ET — and in the past, the messaging from Nvidia, AMD, and others in the graphics space was that the chances of seeing native support for ray tracing engines or for ray tracing to replace rasterization were fairly small.

Real-time ray tracing has historically been treated as a technology that was 4-5 years away — and always would be, apart from the use of specialized accelerators. Of course, history does sometimes validate such technologies, with OLEDs finally on the market and extreme ultraviolet lithography (EUV) tiptoeing towards broad introduction. Then again, OLED remains confined to phones and high-end TVs and displays, while EUV still faces challenges. Nvidia believes the tensor cores aboard the GV100 can be used for denoising and other workloads that should be beneficial in professional contexts as well.

Pricing hasn’t been announced yet, but expect it to run above $5,000; 32GB of HBM2 isn’t going to be cheap, and neither is the full Volta core. This is a card that Nvidia is positioning as bringing important features to multiple new markets in the professional visualization space and it’ll be priced accordingly.

What’ll be particularly interesting for consumers is seeing which features from Volta make it into future consumer cards based on Nvidia’s upcoming consumer GPU architectures. Unfortunately, for now, NV is keeping its lips sealed on that topic.