GTX480 MAY HAVE ‘LAUNCHED’ but the story is far from over. The first derivative part, GF104, is coming, and there is a mysterious new code name on the far horizon.
SemiAccurate has confirmed that the GF104 did tape out a few months ago. If all goes well, there is about a six month lag between tapeout and chips to sell. With its big brother, the ill-fated GF100, Nvidia is at 10 months and slipping, so don’t count on GF104 to be on time. GTX480 might have been ‘launched’, but cards are still not for sale.
For GF104, likely called the GTS450 or GTS460, word on the street is that it will be cut in half from 512 theoretical shaders to 256. The real open question is what will the ‘uncore’ or non-shader section of the chip end up as, half of GF100 or the full uncore? GF100 is horribly weak in pixel ops. It has only 64 ROPs (Render Output Units), and that lack of pixel power shows up at higher resolutions.
If Nvidia cuts GF100 cleanly in half to get GF104, it will also suffer from the same handicap. If the GF104 has more than half the GF100’s uncore, the price is paid in die area and power instead of performance. The GF104 looks to be about the size of AMD’s Cypress HD5870, consume around the same power, and sit between the HD5770 and HD5830 in performance.
The short story is that when you have a chip that is too large, too hot, and badly underperforming, cutting it in half will still put you at the same ratio disadvantage against your competition’s ‘half’ chip. GF104 is unlikely to change the game, either in performance or revenues. Nvidia will likely suffer from more than its current 60 percent die size disadvantage compared to the analogous ATI GPU. The more things change….
Getting back to the GF100, sources at Nvidia are confirming that the yields are far less than 20 percent on both of the variants, GTX480 and GTX470, combined. This means that the silicon for the GPU, before packaging and testing, costs at least $250 for each part. Once you add all the components and make the card, there is no way the GTX470 and GTX480 can make a profit, given what they are being sold for.
If you are queued up for the second batch, don’t expect there to be one. If Nvidia is losing money on each board it sells, trying to make it up in volume is not a sane way to proceed. Expect Nvidia to once again pretend that the cards are not EOL’d and trickle the stock it puts aside out over the next few months. If this sounds eerily familiar, that is because Nvidia is doing the exact same thing with the GTX285. It is keeping up appearances while its partners get no stock and suffer the financial consequences.
Last up is a new code name for the far future, GF119. It is likely a next gen ‘Fermi II’ part, and if Nvidia keeps its naming conventions, a low end one at that. There hasn’t been an x9 part in the last three generations of GPUs, so this might be an attempt to do something altogether new. People briefed on Nvidia’s long term plans say that this is one to keep an eye on, so SemiAccurate will do exactly that. Stay tuned.S|A
Latest posts by Charlie Demerjian (see all)
- Another Qualcomm server core breaks cover - Apr 24, 2015
- Ubuntu strips a phone OS to the Core - Apr 21, 2015
- What does Qualcomm’s server SoC look like - Apr 15, 2015
- How does Qualcomm’s SenseID fingerprint scanner work? - Apr 9, 2015
- How fast is Qualcomm’s 64-bit Kryo server core? - Apr 7, 2015