NVIDIA Fermi core misses expected clock targets?
Semiaccurate are reporting that the Fermi clock speeds are way off what NVIDIA themselves had hoped for.
As you'll remember us reporting a couple of days ago, today (Tuesday 17/11/09) is the day NVIDIA are going to demonstrate the Tesla variant of the Fermi at the SC 09 show. Semiaccurate are reporting that even these wont be the fully fledged boards.
Semiaccurate say :
Sure, Nvidia will have one or two 'puppies', the firm's code for Fermi boards, sans silicon, hacked up, and stuck together with the finest glue sticks on hand. It will even have one or two *GASP* with silicon in them! The problem? The silicon blows.
There have been many rumours, including on this site, about the ability of NVIDIA to produce 40nm chips in the volumes needed. ATI are finding it difficult to produce sufficient 5xxx series cards and that was much further advanced down the development road than Fermi.
Semiaccurate again :
Nvidia can't yield a wet napkin with 40nm engravings on it right now. Tomorrow's GT240 is almost a year late, underpowered, under-performing, and too expensive to be taken seriously. At $99, it gets eaten alive by ATI's last generation, and thoroughly abused by the current one. If Nvidia is four months late with an ~130mm^2 chip on TSMC 40nm, does it really expect us to believe it can make a 530++mm^2 chip in anything approaching volume? Granted, it will make more than seven, but not enough to make a real product launch. Expect more paper and pomp than actual shipping product for this 'hard launch'.
However, it doesn't matter if the product is fantastic. We all love hardware, NVIDIA have given us some of the most advanced GPUs for many years, so if it is mind-bendingly brilliant then any delay is worthwhile. Unfortunately it appears that even the engineering samples are vastly missing their anticipated clock speeds, which could really put the pressure on the architecture itself to produce the goods.
Semiaccurate one final time :
The worst part is the clocks. No, really. The numbers we have been hearing for GT300/Fermi for the past few months were 768 GigaFLOPS for double-precision floating-point and 1.5 TeraFLOPS in single-precision floating-point, compared to the 544 GigaFLOPS double-precision floating-point and 2.7 TeraFLOPS in single-precision floating-point for the ATI HD5870 or Cypress chip. That might not seem like a problem until you look at what Nvidia said in its presentation, that is, 520 to 630 GigaFLOPS in double-precision floating-point. A quick trip to the calculator says 630/768 = 0.82, or nearly a 20% clock miss.
NVIDIA themselves have confirmed that the expected release date is in the first quarter of 2010. Read their press release here
Do you find the potential lack of performance worrying? Have you switched to ATI or are you holding out for the Fermi? Discuss this in our forums
, we'd love to hear what you think.