Ever since it became clear that Nvidia’s Turing would be built on 12nm rather than 7nm, there’ve been questions about when Nvidia would introduce a 7nm GPU. AMD was first to the new node with its professional Vega GPU and intends to launch new Navi hardware in 2019 on the node as well. Nvidia, in contrast, has been quiet about its plans for 7nm and a next-generation node. Now, there are fresh rumors that Nvidia will tap Samsung for 7nm and build a GPU on that company’s process node for 2020.
This could be true. In fact, with just two foundries left (apart from Intel) with leading-edge manufacturing capabilities, we can guarantee Nvidia is using either TSMC or Samsung. Shocking prediction, I know.
It’s unclear how to parse the tea leaves on this one. On the “believable” side of the equation, there’s the fact that Samsung seems to be hurting for major EUV customers. IBM is the only major declared customer for its latest process node that we’re aware of. Picking up Nvidia would be a significant win for the company. And while Samsung serves as its own manufacturer and can, therefore, count on building its own parts in-house to keep fabs occupied, having customers still allows it to spread development costs. It’s not clear that EUV really gets Nvidia anything in this context — there haven’t been any public analyses of Samsung’s 7nm + EUV process to compare against TSMC’s non-EUV 7nm process, at least not yet. But it could be that there are specific benefits Nvidia believes it can gain.
On the “less believable” side of the equation, TSMC hit 7nm first and the first foundry to the node tends to capture the majority of the business. TSMC and Nvidia have a long-standing partnership and TSMC’s 7nm node is rumored to be running below capacity, which would give the company more reason to offer Nvidia sweet terms for manufacturing. On the other hand, one can flip that last point around and theorize that TSMC might be running below capacity on 7nm (assuming the rumor is even accurate) because Nvidia isn’t going to tap them for 7nm manufacturing.
There have been rumors before that Samsung and Nvidia were partnering up on 14nm — repeated ones, as you can see from the various dates on these stories. Ultimately, nothing much seems to have come of them. There have also been previous rumors that Nvidia and TSMC were already signed up with one another on 7nm. Part of what makes these rumors common is that they may all be accurate to some degree.
A company like Nvidia might choose to run test silicon at multiple foundries to see how well the node appears to be suited to their own designs. It might choose to bring up the same design at two different foundries to make certain of having an adequate supply. It might change foundry partners altogether if the first company proved incapable of building the design at all. Second-sourcing is complicated and completely moving a design to a different foundry is a worst-case scenario for a company, but we’ve seen both happen in the past few years. Apple used both Samsung and TSMC for its 14nm SoCs and AMD had to cancel the Krishna and Wichita APUs it intended to build at GF (these became Kabini and Temash as eventually constructed by TSMC). And, of course, a company might choose to build some designs with one foundry and some designs with another, depending on the specific characteristics of the nodes in question.
Finally, companies sometimes deliberately occlude aspects of their manufacturing plans. Take the GlobalFoundries’ announcement that it would cancel its entire 7nm development plan. AMD and GF had previously made multiple public statements about their intent to develop and deploy 7nm technology. Yet GF’s decision to cancel its 7nm node had no acknowledged public impact on AMD’s 7nm Ryzen timeline. That’s exactly the opposite of what happened when AMD canceled the Krishna and Wichita APUs it intended to build at GlobalFoundries back in 2011. Moving those chip designs to TSMC cost AMD a great deal of time. As we wrote then (on Nov. 22, 2011): “a true 28nm replacement for Krishna/Wichita is likely at least 18 months away.”
AMD launched Kabini on May 23, 2013 — exactly 18 months later. It announced no equivalent delay for Rome or next-gen Ryzen. Clearly, the company had reason to believe it would partner with TSMC over GF for 7nm long before GlobalFoundries made its own announcement.
As for the specific rumor that Nvidia won’t launch Ampere until 2020, that could be directly related to AMD’s own plans. I’ve said before that I think it’s unlikely that Nvidia will allow AMD to claim a new node for 6-12 months without launching some kind of response of its own. I still think this is true, but since we don’t know when Navi will launch or precisely which segments it will target, there’s play in the final numbers. It’s not difficult to imagine a July-September 2019 launch for Navi, followed by a January-March 2020 launch window for Ampere.
If Navi is a midrange play that tops out at $200-$300, Nvidia may not care enough to respond with a new top-to-bottom architectural refresh. If Navi does prove to be competitive against the RTX 2070 at a lower price point, Nvidia could simply respond by cutting Turing prices while holding a next-generation architecture in reserve precisely so it can re-establish the higher prices it so obviously finds preferable. This is particularly true if Navi lacks GPU ray tracing and Nvidia can convince consumers that ray tracing is a technology worth investing in.
A review of Nvidia’s conference calls and presentations going back to its Q2 conference call yields no information on its 7nm rollout. The one thing we can say is that Turing is virtually certain to be a shorter cycle than Pascal was. Pascal set a record for the longest-lived high-end GPU family in Nvidia history, anchoring the top of Nvidia’s product stack from May 27, 2016 to September 19, 2018, or just shy of 28 months. It even beat AMD’s Fury X for longevity (the previous record-holder ran from June 24, 2015 to August 13, 2017). For Turing to match its predecessor, Nvidia would have to have to be willing to push Ampere all the way back to January 2021. That seems very unlikely, no matter what AMD or Intel do.