SemiAccurate moles are saying that as of the latest roadmap update handed out about two weeks ago, there is a GeForce based on GK110. This was not the case on any prior roadmap, and it has some very interesting implications. If you are thinking that I mean upsetting the GPU performance ladder, that has nothing to do with it, GK110 is going to underperform expectations on the desktop because it is an HPC chip, not a gaming chip.
The interesting question is why Nvidia would put this beast out for the desktop at all, it is unsuited for that role. Like the Fermi based GeForce GTX465, the answer is simple, yield. Nvidia is the only company having problems with TSMC’s 28nm process, everyone else is at expectations or better. While things have undoubtedly been getting better since the Kepler launch, Nvidia is undoubtedly behind the curve. Again.
The GK110 architecture is not a good candidate for the desktop because the main advances over GK104 are massive DP FP power that consumers will never use and putting back in HPC features that were removed to save power. Some of the HPC tech will show up in consumer benchmarks, but nothing close to commensurate with the die area they consume. The transistors added to GK110 also burn power for very little added gain. Performance per watt is going to suffer badly.
The transistor count of GK110 is about twice that of GK104, so the die area of GK110 is going to be in the ballpark of twice GK104’s roughly 300mm^2. This is comparable to the biggest GPUs that the company has made in the past. Given the problems that Nvidia had with the 40nm Fermi, coupled with their lingering 28nm woes, you have to wonder what yields will be like on GK110. If you are thinking lower than hoped for in Santa Clara, you would almost assuredly be correct.
GK110 is eagerly anticipated in the HPC world, and Nvidia would like to supply those customers with as many high margin Teslas as they can. If yields are low, there are two paths for a chipmaker to take, run lots of wafers to get perfect chips to sell to the “whales” or cut down the specs of the card until it fits the yields coming out of the fab. The latter strategy was used for the Fermi based Teslas, but this time the competition is much fiercer. Worse yet for Nvidia, the competition can actually make their products at high yields for a lower cost. Castrating GK100 is unlikely to hold the line in 2013.
So that means Nvidia will likely have to put out the GK110 based Tesla at the highest possible spec, the top bin goes to the HPC teams this time. If yields are high, this works out well, you can cherry pick parts for HPC and still have enough for credible mainstream consumer use. If the stars all align, you might even have a better part than you expected for HPC too.
If yields are low however, you end up with boatloads of partially working GPUs that have no way of being sold. Given that Nvidia will be getting roughly 100 die candidates per $6000 wafer, so you can do the math on costs with whatever yields you think are appropriate. No matter where the yields end up, there will be a lot of expensive scrap.
The choice in the end is keychains or roll out a reassuringly expensive GPU that no sane buyer would want. A bit faster for a lot more power and money is not a surefire win, but there is a pool of buyers with more money than sense. If people will buy, why not sell? In the end, Nvidia is launching a GK110 consumer part. When the final SKUs are released, the biggest news will not be about gaming but where the company is on the 28nm learning curve.S|A