Nvidia has been busily blaming TSMC for their many supposed failings on 28nm, but SemiAccurate has learned what is really going on. The short story is that the finger pointing around the Kepler launch problems should be at Nvidia, not TSMC.
You might recall that all TSMC 28nm customers are complaining about tight wafer starts, but other than one little ‘whoopsie’ that still hasn’t been adequately explained, all now seems well in Hsinchu. Customers, other than Nvidia, don’t seem to be complaining about yields, quite the contrary, most are quite direct in stating that they don’t have a problem with 28nm yields. Please be aware that yields are not directly related to wafer start availability. By their own admission, Nvidia is having yield problems, and big ones at that. Any wafer start shortages are problems on top of the horrendous yields.
Their flagship GK104/GTX680 GPU has been shipping for a month or so, and availability is still non-existent, completely sold out at any credible retailer. Moreover SemiAccurate sources say that the initial shipments of GTX680s did not hit 1000 worldwide, and to date, less than 10,000 have been shipped in total. Compare and contrast that to AMD’s Tahiti/HD7970/HD7950 chips which had an initial shipment of over 10,000 units, a second shipment before launch, and many more shipments to date. This is in spite of AMD producing a larger chip on a process in a much more immature state at time of shipment. The two smaller GPUs, Pitcairn and Cape Verde with different mask sets from Tahiti, have shipped proportionately greater volumes on top of that, AMD is unquestionably not having problems with yield.
Still, non-existent supply is a symptom, not an explanation, and Nvidia keeps insisting that the shortages are a perception problem, not a supply issue. This however is directly contradicted by both their statements and the leaks from green haired moles. The one key piece of the puzzle that SemiAccurate moles recently managed to confirm was that the problem is indeed Nvidia’s fault, it is a design problem.
This was confirmed when another source and said Nvidia is looking to redo their masks. He didn’t know where in the process things stood, but did say a base layer change is likely, basically the worst case scenario. If Nvidia has to redo this, it is a two quarter process before yields improve, so Q3 best case for yield improvements if they started things this year. SemiAccurate does not currently know where in the decision making process things currently stand though, and Nvidia may just tough it out until the GK114 update makes it a moot point.
As things stand, several questions remain open. When was the defect found? How bad are yields? Does it affect the other 28nm die, the GK106/GK107? Can Nvidia supply OEMs the mobile GPUs they promised? Can they make any Kepler part profitably at this point? One thing you can say for sure is that the problems Nvidia is having on 28nm are self-inflicted, TSMC is absolutely 100% blame free in this instance.S|A
Latest posts by Charlie Demerjian (see all)
- Qualcomm ups Spectra ISP to 2.0 - Aug 15, 2017
- Coffee Lake is going to impact Intel’s margins - Aug 10, 2017
- SemiAccurate digs up Intel Coffee Lake specs - Aug 9, 2017
- Everspin hits the 1Gb milestone with new 28nm MRAMs - Aug 8, 2017
- Intel Xpoint/Optane still doesn’t work right - Aug 7, 2017