As we step into the latter half of the first-quarter release time frame for Nvidia's GF100, Charlie Demerjian of SemiAccurate has published another scathing article that labels the upcoming graphics processor as "unworkable, unmanufacturable, and unfixable."
According to Demerjian, Nvidia is facing silicon yields in the single digits with the current GF100 spin—and that's with the GPU clocked down to 600MHz with 64 of its 512 stream processors disabled. After three spins, the GF100 is allegedly not doing much better than the first iteration Nvidia got back in September. (Word is that spin topped out at 500MHz with similarly low yields.)
The problem, in Demerjian's view, is that Nvidia failed to do its homework by working around issues with TSMC's 40-nm process using smaller, previous-generation GPUs first, then carrying the workarounds over to the GF100. While AMD was making "salable quantities" of its 40-nm RV740 in April of last year, he claims, Nvidia only managed to ship its 40-nm G216 and G218 GPUs to PC vendors in August, and the cards only hit retail in October. Nvidia had trouble enough getting those low-end GPUs with footprints around 100 mm² out the door, and then, it immediately turned around and tried to get a 550 mm² design through. AMD, on the other hand, had time to implement workarounds learned from the RV740 in its Evergreen series of GPUs.
The result, he claims, is lower-than-expected performance, very high power consumption, and extremely poor yields for the GF100. If a single GF100 wafer costs Nvidia $5,000 and only contains 10 usable GPUs, Demerjian points out, then each chip will cost the company $500. That's not counting the bill of materials for the rest of the card, either. By contrast, the article suggests AMD's high-end Cypress GPUs only cost around $50 a pop.
To do things right, Demerjian reckons Nvidia would have to retool the GF100, do a "base layer respin," and follow up with "at least one metal spin on top of that." That process would take six months, and even if Nvidia got started today, the re-architected GF100 would show up mere months before the arrival of the first 28-nm graphics processors. In the end, the only Nvidia DirectX 11 GPUs he expects we'll see before the end of this year could be limited numbers of GF100 cards, all sold at a loss. (GF100 derivatives reportedly "haven't taped out yet.")
Demerjian is known for mercilessly slamming Nvidia, of course, so you may want to take this report with a grain of salt. Still, Nvidia could be in trouble if the situation is even half as bad as he claims.