Last year's GTC event in San Jose, California saw the unveil of nVidia's architecture following Maxwell: Pascal. We wrote about Pascal at the time, but very little was revealed about the new architecture. This year's GTC keynote presentation by nVidia CEO Jen-Hsun Huang revisited Pascal architecture and the nVidia GPU roadmap through 2018.
Sometime within the next eighteen months, nVidia hopes to launch its Pascal architecture, which will lead into Volta closer to 2018. Pascal will launch alongside NVLink compatibility, nVidia's higher-bandwidth PCI-e alternative for high-precision computing. Note well that this is not a product initially intended for the gaming audience; NVLink and Pascal -- strictly in their currently-unveiled form -- appear to be more research-targeted than gaming-targeted. Most architectures work their way down from research and enterprise into the gaming market, so it's still worth talking about.
Pascal's major differences, as discussed last year, are its introduction of 3D memory (a trend in the memory industry) and mixed precision computing. The keynote informed onlookers that Pascal would increase capacity by approximately 2.7x, increase bandwidth by 3x (or 6x on the 3D memory side), and grow FP16 mixed-precision COMPUTE performance approximately 4x. Using "CEO math," Jen-Hsun Huang indicated a rough "10x increase" in deep learning performance of Pascal over Maxwell.
Pascal's benefits are clear for deep learning and research tasks. We're yet unsure of how the architecture will work its way into the gaming world, though a smaller scale form of Pascal will likely show itself in gaming GPUs in the semi-distant future.
No discussion on Volta was offered during the keynote.
- Steve "Lelldorianx" Burke.