Miracles happen, GT300 tapes out!

Warning: flying pigs and huge chips

LOOK, OUT THE window, a squadron of flying pigs! No, really, they do fly, there are frozen lakes of fire being reported all over, and Nvidia’s GT300 has finally taped out.

Our sources tell us that the GT300 is beginning it’s 7 or 8 week long process through the very expensive machinery at TSMC. We said the target was a mid-July tapeout, and if you are very charitable with ‘mid-‘, Nvidia hit it. A hair under a week ago, the chip had not taped out.

Now it is time to peer into the giant crystal ball of semiconductor scheduling. Lets assume that the chip taped out on work week 28, more or less July 15. TSMC runs 6-8 weeks for a ‘hot lot’, but our moles say currently it is much closer to 8 than 6. Add in just over 2 weeks for initial card build, bring-up, testing, debug, and making any fixes. Lets just call it 10 weeks, October 1 or so.

If everything goes perfectly, and the GT300 is literally bug free, or possibly having only minor bugs that Nvidia is totally confident can be fixed without causing any more problems, then they can start running production silicon.

From there, TSMC takes 10-12 weeks to run 40nm wafers, there is a lot to do on those parts. If we are optimistic and say 10 weeks, and add another two for the initial board manufacture and shipment, we will be at a total of 22 weeks out. This means if everything goes perfectly, not so much as a hiccup, we are looking at mid-December for the first syphilitic trickle of parts.

The chances of this happening are somewhere around zero though. Nvidia can’t make GT215s in volume, and that chip is about 1/4 the size of GT300. Toss in that this is going to be their first GDDR5 part, their first DX11 part, and their first part on a new and untested architecture, and you are likely to have at least two spins.

Update: It is not their first GDDR5 part, they have two out. We even wrote it up weeks ago here. Duh.

Each spin is the same 8 weeks as the last hot lot. Adding one puts production silicon out of the door, with the previously alluded to trickle, in late February. A second spin puts it into Q2. If there is anything more, well mid-year is very likely. Smart money is on at least two spins.

That said, not having correct or volume manufacturable silicon never stopped Nvidia from having a launch. If Dear Leader wants a launch in 2009, he can have a launch in 2009, and the press will get $15,000 hand built cards that use sub-10% yield parts. It will be ‘launched’ though, but calling that soft is being far too kind.

It will be interesting to see what Nvidia does to counter the ATI launches, they can likely show something off about the time ATI Evergreen parts flood the market, but it won’t be production ready silicon. Expect a seriously disingenuous and desperate spin campaign with slides that bend any reasonable definition of honesty.

Back to the chip itself. Our moles say it is 23mm * 23mm, or about 530mm^2, and considering they told us that GT200 was 24*24 months before it came out, we will take their word for it. This puts GT300 almost dead on even between GT200 on 65nm and GT200 on 55nm. If you recall all those charts about alleged GT300 sizes and specs floating months ago, we will say the same thing we said then, complete fabrications.

In any case, Nvidia is still having massive problems getting any chip out on TSMC’s 40nm process. The two that are ‘out’ now aren’t really available in any quantity, amazing since they are sub-100mm^2 parts. The 12*12mm or so GT215 isn’t slated to come out until December, so what chance is there of something 4x that large making it?

To add more data to the fire, the initial GT200 was built on a known, understood and proven TSMC 65nm process. It had the bugs worked out long before NV tried their hand at the 576mm^2 GT200. Our sources tell us that the yield for that was in the low 60s when they pulled the trigger on production, far higher than they can hope for with GT300.

That brings us to the question of the moment, what is the shortest time between tapeout availability for any modern GPU that was a new architecture, on a new process? G80 wasn’t on a new process, R670 wasn’t really a new architecture. Post the answer in the forum, the first one to get it right wins nothing. Two nothings if you include tapeout dates.

In the end, there won’t be any GT300s this year, and early next year is really iffy. Press stunts aside, it is vaporware for 2009, but it has, finally taped out.S|A

The following two tabs change content below.

Charlie Demerjian

Roving engine of chaos and snide remarks at SemiAccurate
Charlie Demerjian is the founder of Stone Arch Networking Services and SemiAccurate.com. SemiAccurate.com is a technology news site; addressing hardware design, software selection, customization, securing and maintenance, with over one million views per month. He is a technologist and analyst specializing in semiconductors, system and network architecture. As head writer of SemiAccurate.com, he regularly advises writers, analysts, and industry executives on technical matters and long lead industry trends. Charlie is also available through Guidepoint and Mosaic. FullyAccurate