Nvidia might be forced to re-engineer Maxwell for 28nm, if a new report holds true.
Relations between Taiwan Semiconductor Manufacturing Company and Nvidia are likely at an all time low, as a new report says that TSMC isn’t able to deliver the next two Maxwell-based followups for its GPU line, the GM204 and GM206, at 20nm forcing Nvidia to do a last minute reengineer at a 28nm process node.
According to a report from 3D Center, the GM204 would be taping out in late-April for a December product release and the GM206 would be taping out around the same time but would not be released in products until January 2015. The GM204 is the successor to the GK104, while the GM206 is the successor to its mid-range GK106 silicon.
This report from 3D Center follows a scoop from Sweclockers earlier this month citing sources in the Taiwanese supply chain which said that 20nm Maxwell would be delayed possibly as late as next year due to TSMC not being ready to handle a large scale 20nm production order.
Another alternative theory presented in the forum posts that 3D Center’s report sources from is that Nvidia has figured out a way to engineer the intended performance gains Maxwell would provide on 20nm on the existing 28nm process. But this theory seems less likely when one looks at the history of Nvidia’s relationship with TSMC.
An icy relationship
Most business relationships within the technology industry are governed by strict non-disclosure agreements. ODMs, for example, must maintain airtight secrecy in their organizations as their entire business model depends on being able to manufacture a company’s product without leaking the specifications to an outside party or even to one of its own divisions that’s responsible for making a competitors products.
So it would be highly surprising to see a company break the relationship of mutual silence with an ODM and criticize it, but that’s exactly what Nvidia did to TSMC at the 2012 International Trade Partner Conference.According to slides obtained by ExtremeTech from Nvidia’s presentation (the conference was only open to industry members and not the press), Nvidia presented a stinging critique of TSMC’s struggle with adapting its production techniques and its corporate culture.
Though the difficulties in dealing with TSMC’s corporate culture isn’t surprising to anyone who’s dealt with the ever-challenging “cult of the Laoban” managerial style present in Taiwanese corporate institutions, the real takedown came from the frank disclosure that Nvidia disclosed — publicly — that it thought TSMC wasn’t up to the challenge in giving Nvidia the required price-performance ratio at 20nm. The added costs when dealing complexity of the process node just wasn’t worth it in Nvidia’s opinion, but what else could Nvidia do but try and shame them into doing better? It really was an incredible and risky move on Nvidia’s part.
What to do about Maxwell?
Putting the report from 3DCenter and SweClocker in context, it looks like a delay because of process node issues is probable. It’s been in the cards since 2012. The real question is what is the alternative to fabricating with TSMC? Would Nvidia take the risk and jump to GlobalFoundries — a highly unlikely option considering the company’s history with AMD. Or will it look for a new partner?