Nvidia has finally unveiled its new GPU architecture - codenamed Turing - with three products designed primarily for the professional market, while teasing the upcoming reveal of next-gen gaming graphics cards - where they may be a couple of surprises in store.
At this year's Siggraph, Nvidia boss Jen-Hsun Huang unveiled three Nvidia Quadro RTX cards - the RTX 8000, RTX 6000 and RTX 5000 - built from two different Turing GPUs, with 4608 and 3072 CUDA cores respectively. The design appears to be a significant evolution from the Volta architecture seen in the Titan V, retaining the deep-learning AI Tensor cores, and doubling down on dedicated ray-tracing hardware.
As expected, there's significant change in the nature and performance of the attached VRAM too, with both of the new processors paired with new GDDR6 memory. The 'big chip' offering delivers 672GB/s of bandwidth vs the 448GB/s for the second-tier product. For comparison's sake, the current GTX 1080 Ti delivers a higher 484GB/s of bandwidth. Although not confirmed officially, it's almost certain that the higher performance chip runs on a 384-bit memory interface, up against a 256-bit bus on its stablemate in the RTX 5000.
Interestingly, on the larger chip, the CUDA core count has actually dropped by 512 cores compared to the Volta GPU found in the Titan V, and if we are to assume that the RTX 5000 processor outperforms the GP102 found in Titan XP and GTX 1080 Ti, the reduction in CUDA cores there is even more significant - 3072 vs a max 3840 on Titan Xp. However, Turing isn't a smaller-scale evolution in design in the way that Pascal was - it's a complete revamp, with the cache architecture alone seeing a 2x increase in bandwidth - a smaller amount of cores should still be able to deliver more performance.
While Nvidia isn't directly discussing the gaming products - yet - we can safely assume that a version of the RTX5000 product will form the basis of the new gaming flagship. By extension, the larger RTX 8000/6000 processor could take the form of a new Titan/Ti product further on down the line. However, based on yesterday's reveal, the size of this chip could make widescale production challenging - it's an enormous 754mm2, up from the still-large 471mm2 of the GP102 found in Titan Xp and GTX 1080 Ti.
There are some further pointers about what may appear in the new GeForce line, beyond some of the hints dropped in the video above. For example, there's a USB Type-C connection on the new products based on the new VirtualLink protocol. The aim here is to cut down on unwieldy VR headset cabling by combining video, power and USB connections into a single cable.
Returning to the matter of the teaser video, the sequence kicks off with visual hints suggesting a new cooler design that looks similar to the new Quadro RTX cards, before bombarding us with hints via chat room aliases including RoyTeX, Not_11, Mac-20 and Eight-Tee - tying into rumours that the new flagship is named RTX 2080, as opposed to GTX 1180. There are even GPS coordinates in the video - 50.968495,7.014026 - which resolve to a location on the edge of Cologne, the city where Nvidia is hosting its 'GeForce Gaming Celebration' on Monday next week.
All of the signs point to a new gaming GPU reveal then, and if RTX technology is core to the product, we're looking at a graphics card product that's something a little different from the norm. Typically, a new GPU is all about the percentage uptick in performance compared to the current line-up - but what if the new cards actually deliver something completely new to PC gaming? And if Nvidia is dedicating so much silicon area to RTX technology, just how much support is lined up in upcoming software? We should find out next week, where we'll bring you the news as soon as we can.