IT LOOKS LIKE Fermi A2 silicon has finally taped out, so the timetables are a little firmer once again. There is no chance of a real launch in 2009, making the chip a shining example of Nvidia’s engineering mire. (NASDAQ:NVDA)
Lets recap a bit. We said that Fermi, then called GT300, taped out on about Work Week 28 (WW28), and it did. We said that silicon was due back in 6-8 weeks, and cards could possibly be shown publicly on Oct 1. We admit that we overestimated Nvidia’s ability to engineer its way out of a wet paper bag with a map, flashlight and a bunch of wood screws here.
It hasn’t shown a card yet at all, yields are miserable, but it did in fact get silicon back on either WW35 or 36, which is right where we said it would be, almost to the day. The fact that yields were a joke, coupled with ‘puppy’ inflicted own goals, made things downright laughable for Dear Leader and company. Nvidia didn’t have enough working dies to do the testing it needed, much less show some off for PR, so it faked that.
Back to the chips. Normally the debug and respin process is about two weeks or so, a marker that should have been passed before not-Nvision. As of mid-October, we heard that NV didn’t know what the problem was, and that it was going down the metal stack to desperately try and figure it out. People inside honestly honest green said things were rather desperate.
The latest word was that the chip was set for a WW42 tapeout, or was imminent that week. Let’s give Nvidia the benefit of the doubt that it did tape out, something anecdotally confirmed by Fudo saying that the chip will be out in early December. If by ‘out’ he means A2 silicon samples, then it will in fact be out in December. If you use a definition of ‘out’ grounded in the reality that humans occupy, then no chance.
Assuming that Nvidia parked a few wafers to speed up the next hot lot, it could indeed have a few A2 chips in late November, and boards to show a week or so later. The go or no-go decision could be made on December 1 if all goes perfectly.
From there, if the risk wafers did not need to be scrapped, you are about six weeks from production silicon, best case. Add another two weeks for boards, and you are into February. Given Dear Leader has scheduled a press conference at CES on January 6th, that should give you a good idea of the public timing. For those curious, although Nvidia seems to have forgotten to send us an invitation to their yawner, it will be at noon at the Venetian Hotel on January 6.
Anyway, if all goes perfectly, we are looking at February for the start of real quantities. There will be A2 silicon before that, but nothing in real quantities. Anyone who says otherwise has ulterior motives or doesn’t understand how the industry works.
During not-Nvision/GDC, Nvidia was telling people who mattered and AIBs not to expect Fermi until March. Internally it was saying May, but the AIBs were not told that. About the A2 tapeout time, Nvidia’s AIB messaging was changed to April or May.
It would be safe to read into this that the A2 stepping is not going to cut it, and an A3 spin is on the cards. Eight weeks added to early February gets you into March, so that lines up nicely with what Nvidia is telling people.
Another bit of anecdotal evidence is that there is no sign of the other four GT300 variants taping out. Those are usually kept in house until the first chip is fully baked, and the fixes are backported. If A2 would have done the trick, there would have been much more movement at TSMC on the variants, and there does not seem to be.
To wrap it all up, A2 is out, but it took about four times as long as it should have. A3 seems very likely, and the chance of anything more than a PR stunt launch in 2009 is zero. Don’t believe the hype, Q1 is best case. When you don’t have product, spin.S|A
Latest posts by Charlie Demerjian (see all)
- Intel unleashes more Kaby Lake SKUs on the yearning public - Jan 4, 2017
- Qualcomm opens up a bit more on the 10nm Snapdragon 835 SoC - Jan 3, 2017
- AMD’s Freesync 2 changes the display game - Jan 3, 2017
- Coffee Lake points to issues with Intel’s 10nm process - Dec 28, 2016
- Intel goes all Pokemon with code names – really - Dec 28, 2016