Nvidia’s Pascal GPU architecture has been one of Nvidia’s longest-lived product lines in the past 20 years. It further refined and improved the technologies Nvidia introduced with Maxwell, along with making some architectural improvements of its own, and it competes extremely well against AMD’s Polaris and Vega product lines. But Pascal is also old for a high-end GPU architecture, particularly one that hasn’t undergone a major SKU shakeout since launch. While Kepler beats it in age (GTX 680 debuted in March 2012, with the GTX 980 arriving in September 2014), Nvidia eventually replaced its top-end GTX 680 SKU with the GTX 780, which offered significantly boosted performance, as well as various improvements throughout the 7xx stack. Beyond launching the GTX 1080 Ti almost a year ago and the 1070 Ti late last year, Nvidia hasn’t made any changes to the Pascal family.
For the past few weeks we’ve seen rumors popping up that Nvidia would soon introduce a new GPU family, possibly at the GTC or GDC conferences this year. We haven’t discussed the topic because all our attempts to confirm those rumors pointed in the opposite direction. Now, THG has chimed in, indicating that their rumor mill suggests we won’t see new consumer GPUs based on the upcoming Turing architectures in the next few months.
Nvidia’s Volta uses HBM2, but there’s no sign NV will adopt it for its consumer cards. GDDR6 is the expected memory standard for next-gen high-end cards.
Nvidia is supposedly prepping new cards based on Volta for non-gaming workloads with a new compute architecture, supposedly nicknamed Ampere, coming in the near-future as well. The consumer follow-up to Pascal is reportedly nicknamed “Turing,” though we’ve seen some publications arguing that Ampere is the consumer part and Tesla the compute-oriented GPU. Ampere as the compute GPU seems to align more closely with “Volta” as a brand name, but we’ve got no inside information on that one way or the other.
It feels a bit odd to say this, but the biggest questions surrounding Nvidia’s next-generation GPU, whatever codename it uses, have relatively little to do with gaming performance. There’s significant concern future GPUs could be impossible to find in market if they outperform current Pascal chips at cryptocurrency mining. We already saw this happen with 14nm chips from AMD and Nvidia at the 14/16nm node and nobody wants a sequel with ongoing shortages that could take a year or more to improve.
The question of which node these GPUs will use is also up for some debate. TSMC’s 10nm nodes have, as far as we’re aware, been deployed almost entirely for mobile products. TSMC has also guided that it believes its 10nm products are a short-lived node. This is a similar strategy to what we saw at 20nm, where neither AMD nor Nvidia tapped the node for GPUs and it was replaced by FinFETs fairly quickly. It’s possible, therefore, that Nvidia will target TSMC’s interim 12nm node (think “optimized 16nm”) rather than moving to a brand-new node in 2018.
The other reason Nvidia can afford to take its time in many of these efforts is because AMD’s Vega isn’t really nipping at its proverbial heels. Vega 56’s lead over the GTX 1070 was largely eliminated by the GTX 1070 Ti, and the increased cryptocurrency demand for AMD GPUs has made them less relevant to Nvidia in any case. We covered this in some detail earlier this week, but with AMD GPUs currently seeing more cost inflation than their Nvidia counterparts, Nvidia has even less reason to push for new part introductions. Hopefully this also means they’ve had the time to find ways to prevent cryptocurrency miners from yanking all the desktop GPUs out of the market when new products do inevitably launch, and to ensure a robust supply of product so we don’t see the same supply shortages as happened in 2016.