Sources tell SemiAccurate that the first Kepler chips taped out about three months after the first 28nm Fermi shrink taped out. If you remember when we exclusively told you about the dates on those about a month ago, now there are a few more details to add.
First off are the parts themselves, there are two, GK117 and GK107. GK117 was first, and sources describe it as a hybrid Fermi/Kepler part that doesn’t have a display controller. It isn’t a GPU, and it is, based on the code name, very small, so what it is useful for, if anything, is a mystery. This might just be a test chip with Kepler shaders and Fermi interconnects, but who knows?
GK107 is pretty obviously a small laptop part. If you recall, the GF108 and GF106 were the small and mid-range laptop parts in the Nvidia line, the GK107 looks to be a slightly above miserable performance chip. Where Nvidia hopes to sell it depends on performance, but that market is getting awfully thin because of Llano and Sandy Bridge. Either way, don’t expect much other than a price/performance volume laptop part.
That brings us to the big questions, why not a GK100? Where is the big chip? Why did it not tape out first? The answers are; damn good question, not in Hsinchu, and possibly a big problem. There are two overarching possibilities as to why GK100 is MIA, and neither is good. Normally you tape out the big, difficult chip first, and then dice it when the big one is debugged fully. The idea is that any problems will crop up in a big die, but problems in smaller dies may not be everything you encounter on something 4x as large. Just ask Nvidia engineers about GF100 for more on that.
Small chips are easier to debug than larger chips, and easier to design, verify, make boards for than the bigger parts. Small chips also have vastly higher yield than larger chips of the same architecture as well. Small chips first looks to be a time-to-market strategy and/or a short term profit maximization strategy, there are serious long term scheduling problems with doing it that way because your roadmap suffers.
So that brings us to the technical side of why the small die GK107 first? 28nm yield problems? Could the process be the reason? Could the interconnect be as borked as insiders tell us it is, leading to “Fermi II: The smoke also rises”? Who knows, but not doing the big part first is seriously puzzling, and will probably bite Nvidia in the ass. Luckily, they know how to deal with that kind of issue, and the list of valued partners to blame is undoubtedly being drawn up already.
That gets us to scheduling. The first Nvidia 28nm parts are scheduled to sample around October 1, and they are set to be delivered in volume in late December, call it January 1, 2012. The first Keplers taped out three months after the 28nm Fermi chips, so you are looking at April 1 for Kepler. The small Keplers. Not the big Keplers. If A0 silicon works perfectly.
28nm shrinks are a simple thing to do compared to a new architecture. If Nvidia has to do any spins of GF107, you will need at least 6 weeks for every spin, so April 1 becomes May 15 for A1, July 1 for A2 etc etc. Q1 is unlikely to see Keplers in anything other than PR form.
Moving back to the big Kepler, the one the fanbois want, it is still MIA. It is unlikely to tape out until GK107 is fully debugged, and then add a few weeks for the fixes to be backported and those changes laid out. Because of the big chip/small chip problems, there is likely to be an additional spin or two on top of the GK100 tape out, so more time there. Short story, the Kepler part you are expecting is probably mid-2012 best case.S|A