GT300 delayed till 2010GT215/216/218 MIA as wellBy Charlie DemerjianTuesday, 5 May 2009, 16:25IT LOOKS LIKE Nvidia engineering is getting right back into its old fighting form once again. The GT300 chip is delayed until 2010, and the GT21x parts seem to be MIA.The big news is of course the GT300 delay. A few weeks ago it was promising insiders parts in October, a goal that was laughably unrealistic given that the tapeout was set for June. It seems that realism might be poking it's head through the Santa Clara FUD, er... fog, and NV has told those who need to know that it will be Q1 2010 now. Given its timeliness of late, we think it should append a bold "Maybe...." to that though.We are not sure if it is a delay to the tapeout, or a delay for said fog lifting, but I would bet on tapeout delay. In fact, if I were to shake the magic 8-ball of GPU futures very hard, I would bet that tapeout was delayed, and it is still not being realistic on time to market. That said, as of last week, sources deep inside big green had not gotten their GT300 themed tapeout party hats and related bonus materials (One of these too).That brings us to the GT215/216/218. They were set for about now earlier this year, then nothing. The silence is deafening. In fact, the silence is so deafening that Nvidia is jumping up and down talking about mobile 40nm parts now instead. We hear it is having massive 40nm yield problems, something you simply can't have for your intended high volume parts.So, to change the subject, 40nm is all about mobile now! Ignore the fact that ATI has a very successful 40nm part out in the 4770 - Nvidia is telling OEMs to not position 9800 products against that because it can't match the price. The 40nm products that should compete with it are.... Mobile, did we mention Mobile?We are hearing rumors of an editors day in mid-May, so the usual 2-week time line to introduction puts the 40nm launches at Computex. That means Nvidia won't have a competitive part in the meat of the market for almost all of Q2.But all is not lost, we also hear Jen-Hsun is being tapped for a Hot Chips 21 keynote. That should be a morale booster for the green team, and if minor miracles happen, he might even have early silicon to show off. Given the rest of Nvidia's competitive situation, that will be the (singular) bright spot for Nvidia in the near future. µ
Nvidia's spin borders on the truthAnalysis Q1 financial conference call was amusingBy Charlie DemerjianFriday, 8 May 2009, 22:55THE NVIDIA Q1 financial conference call was the usual mix of overreaching self-congratulations and dodged questions. It sounds good on the surface, but if you know what is really going on, you wonder how they can say what they do and get away with it.The call started off in the usual way, with CEO Jen-Hsun Huang doing the 'we are the greatest in categories we made up' thing. The hype was all about Physx, GPU compute, and mobile. All were said to be doing really well and are - just as they have been every quarter of the past two or three years - expected to explode real soon now.If you recall, two summers ago, GPU compute was supposed to make up 5 per cent of Nvidia's sales by the summer of 2008, according to its Investor Relations propaganda. Summer 2008 rolled by, GPU compute sales were at 0.5 per cent, and no one was talking about the cost of those sales. The company claimed that there were none, that people were getting huge grants to use those products and, well... they just smiled. Intel said much the same about the Second Coming Itanium chip Itanic for years.The same story applies to Tegra. It was supposed to have products using it out about three quarters ago, but now Nvidia is on the verge of almost nearly having products ready. Nvidia's executives won't say who the customers are, but they are huge, trust them, they've never lied about it before. Nope, 100 per cent honesty, just ask them.Moving right along, you learn something new every day. When asked about how Ion was doing, Jen-Hsun reported a huge number, $180 Million. When asked about how that number was derived, check out his response to the question from analyst Rajvindra Gill.RG: "Right, but are you including the 9400M that you are showing to Apple as the... as Ion or is this actually two-thirds of that $180 million design wins with other companies outside of Apple that are using Ion?"J-H H: "We don't distinguish the two... I mean, it's the same processor from our perspective. The... we call it, the platform, Ion, and increasingly people are calling the chip Ion, and so we don't distinguish the two, whatever people want to call it. I mean, when people buy the chipset from us, they call it Ion. When they put it into the box and they brand it from the outside, because Geforce is such a terrific brand and contributes so much to the brand value of the end product, people call it Geforce. Do you see what I'm saying? So people buy the Ion but they sell it as a Geforce."Who knew? Ion is not only a brand now, but it encompasses any Geforce integrated GPU. Wow. If I didn't know any better, I would say that this is a blatant attempt to inflate category numbers using intentionally misleading information. Luckily, the analyst in question is not so easily duped as Jen-Hsun might have hoped.Then they go on to talk about Physx, and tout how it is a huge sales success. According to the conference call, there are all of two (2) games that use it. But one is just eye-candy. Wow, and the pipeline is full, just ask them. But don't ask game reviewers, they don't seem to agree.They don't agree with Nvidia's take on this week's flavor of Ion either. Remember, this box has twice the CPU power of the one in the Revo. Can you say high return rate? Oddly, Nvidia didn't mention those numbers during the conference call.So, as with all self-serving spin, the end result is in the bottom line. Nvidia didn't break everything out where possible, but for one where it is possible - the professional line - sales were flat, quarter over quarter. With desktop sales up 50 per cent, notebooks up 28 per cent, and AMD based chipsets up 94 per cent, zero percent growth in a market that is such a stunning success is... at least odd. One might almost be tempted to say that Jen-Hsun's statements are not backed up by the numbers, but that would be using facts, and Nvidia really hates it when you do that.That brings us to the next thing, margins. Recently, Nvidia's margins were in the mid-40s. This quarter, analysts were expecting mid-30s. They came in at 28.6 per cent GAAP, 30.6 per cent non-GAAP. Whoopsie! It was half explained away, but I don't buy the explanation for a minute.A much better explanation is what we have been hearing from OEMs for a couple of quarters now, that Nvidia seems to be addicted to marketshare and will try to maintain it even at the cost of cash losses. If you have cash to burn, that works out, but when times are tough, it gets a bit more scrutiny. Just ask the game developers who used to get sacks of TWIMTBP cash.In any case, Nvidia's marketshare took a jump in Q1. What it is not telling you is that marketshare growth was mainly due to the company dumping about a million units of 65nm DX9 inventory that it had previously written off.If you do the math, that is more than enough to make up for the marketshare unit gains. What is a bit more troubling is that the margins for discrete GPU sales were not disclosed. Using the numbers that Nvidia did give out, the totals don't add up unless GPU margins were about half of the overall corporate margins. That is bad. Bad bad bad. Your core business should not be struggling to have double-digit margins.Why are the margins so low? It's simple, ATI has a better product that is cheaper to make. AMD can force prices down and still make money. ATI, the division, is making money, but Nvidia is not. But it is growing marketshare! Lose money on every one sold but make it up in volume is not a healthy long-term corporate strategy - it's an old Lucille Ball and Desi Arnez sitcom skit! You can't say we didn't tell you so. And think about this, consumer price levels for GPU parts and everything else are way down since that article was written. GPU prices are down about 30 per cent in just the last two months, for example.However - margins, smargins - Nvidia is the leader in visual computing, at least until it makes up a new category. And to top it off, things are looking good. Once again, Jen-Hsun said that the 40nm ramp was doing well. To use his exact words: "Let's see... the ramp is going fine." Bullsh*t.No really, bullsh*t. He should know better than to say that. The only reason it isn't dishonest is that he can sleaze his way out of what "fine" actually means. Nvidia was supposed to beat ATI out to market with 40nm parts. Nvidia was slated for November, ATI for December.The ATI parts, the RV740 / HD4770 came out in May. The Nvidia GT216 / 218 were supposed to beat them out by a few weeks. Three months ago, the parts were supposed to have been out by this time. Right now, the updated schedule - due to a ramp that is going "fine" - targets June or July. Sources in the Far East say things are looking more like late June, and slipping.Toss in that the GT300, another 40nm part, was delayed 'til 2010, and it makes you wonder what the Nvidia definition of "fine" is. Boils? Cafeteria food? If its 40nm ramp is "fine", so is the mess in Iraq. What I suspect is going on is that Nvidia hasn't been able to get its 40nm chips up to an economically viable yield.The now dead GT212 was a die-shrink of the GT200b from 55nm to 40nm, basically a margin improvement part. It is not to be. The GT214 / 216 / 218s are basically GT200 derivative parts, but we are also hearing that they are shrunken G92s. In any case, the largest of them, the GT214, failed like the GT212. It was reworked with an eye toward actually functioning, and evolved into the GT215.The problem here is that the 214 to 215 evolution took so much die area that it is unlikely to ever be profitable. If you can't make a 130mm^2 chip on 40nm, what hope do you have of doing a 500mm^2 part?Yeah, the new savior chips are the old chips but smaller. Given they are now eight ( months late, best case, and don't have GDDR5, it is unlikely that they will be profitable at competitive price points. Toss in abysmal yields, and the 2009 profit margins are more or less non-existent. Best-case timeframe for a competitive mid-range part, Q2 of 2010.That brings us to the GT300. Jen-Hsun artfully dodged the DX11 Windows 7 question. Why? AMD is on record saying that it will have a DX11 part out for the Windows 7 launch, and our sources back this up. Nvidia will be, best case, almost two quarters late to the party. Let me repeat that, Nvidia will have no parts out for either the Windows 7 launch or Christmas, so it is all ATI this year.How they can spin this as good is beyond us, especially if you know the performance levels of both parts. Nvidia's strategy of GPU compute is going to be impacting its performance this round, hard. Nvidia took a stupid strategy with the GT300, and that is going to bite it.Graphics performance sells graphics cards, GPU compute does not. It may be an adder, but without a base, it goes nowhere. ATI will likely have derivative parts out before Nvidia has its first DX11 part in volume, and ATI's current die area and therefore cost advantage should carry over as well.If all goes well, Nvidia can hope to start competing for marketshare again next spring. Until then, things are only going to get worse for the green team.Nvidia's executives can spin the situation all they want, but all the analysts I talk with understand the credibility gap the company has by now and seems bent on expanding.Nvidia is not competitive, and it has nothing in the pipeline that will fix the problem for the next three to four quarters. Until then, well... things look ugly at best for Nvidia. You can only spin so hard before you throw up. µ
I asked two people at NVIDIA why Fermi is late; NVIDIA's VP of Product Marketing, Ujesh Desai and NVIDIA's VP of GPU Engineering, Jonah Alben. Ujesh responded: because designing GPUs this big is "f@#king hard".
Nvidia's Fermi GPU gets cut backToo hot to handleBy David NealWednesday, 23 December 2009, 14:15ACCORDING TO A REPORT at Semiaccurate, Nvidia's Fermi GPU chip has recently been downsized.Nvidia has had to cut back its Fermi GPU because of the chip's low wafer yields and its high heat output, if the latest reports are to believed.Fermi has long been thought to be too hot to handle and Nvidia has now cut down the number of stream processors to 448 instead of the previous 512, and it has admitted that the GPU chip will be a 225 Watt part. Nvidia's own Fermi information pages still show it as having 512 SPs, but according to a recent PDF this is no longer the case.Semiaccurate thinks Nvidia cut back the number of stream processors either because the Fermi chip can't be produced in volume with less than two major flaws in its SP array or its power draw and heat output were just too much for Nvidia's graphics board partners to deal with, or both. Those reasons do sound plausible.But really, a single GPU chip that draws 225W is simply unheard-of. Let's just put it this way - we wouldn't fancy leaving the thing running on top of some dried newspaper and kindling.This new information makes some of the praise that's still on Nvidia's Fermi information page look a little odd. "Fermi surpasses anything announced by Nvidia's leading GPU competitor (AMD)", said Tom Halfhill, senior analyst and editor at Microprocessor Report. Dave Patterson, the director of the parallel computing research laboratory at UC Berkeley, added, "I believe history will record Fermi as a significant milestone."Maybe not so much a milestone as a gravestone, perhaps. Fermi is starting to seem like a GPU chip too far.What seems certain is that Nvidia's fab partner TSMC is finding it challenging, to say the least, to manufacture the chip, and that it will have to be cooled by a waterblock if not a more exotic cooling method like phase-change refrigeration.Whether this will all work out for the Green Goblin remains to be seen. But at this point it looks as though Nvidia's Fermi GPU chip might be delayed again. µ
TSMC claims to have fixed 40nm yield problemsAbout the same level as 65nm these daysBy Ed BerridgeWednesday, 20 January 2010, 11:47CHIPMAKER TSMC said it has improved yield rates on its 40nm manufacturing process and everything is on a par with the older 65nm process.According to Digitimes, Mark Liu, senior VP of operations at Taiwan Semiconductor Manufacturing Company (TSMC) said at a company event that the chamber matching problems that had blighed yield rates for the last quarter have now been fixed.He did not say much else, but the low yields have resulted in problems for several of TSMC's customers and could mean that some will have to delay product launches while the chipmaker clears up its backlog.TSMC opened a new factory building named Phase 5, which is part of the company's Fab 12 located at the Hsinchu Science Park in Taiwan. Once Phase 5 is operational it will be used for volume production of 28nm products in the third quarter of 2010.The outfit is also planning to build Phase 6 of Fab 12, which mainly will be used as its 22nm production base. Liu said all this building is needed to keep up with demand for chips in the coming quarters. µ
WITH ANOTHER LAUNCH of the Nvidia GT300 Fermi GF100 GTX480 upon us, it is time for an update on the status of that wayward part. Production parts have been coming back from TSMC for several weeks now, and the outlook for them is grim.We first got word that production A3 GF100s were back in Santa Clara near the end of January. When enough of them rolled off the line to characterize the silicon, we hear there were no parties in Santa Clara. For those not interested in the 'why' side of things, the short answer is that the top bin as it stands now is about 600MHz for the half hot clock, and 1200MHz for the hot clock, and the initial top part will have 448 shaders. On top of that, the fab wafer yields are still in single digit percentages.That said, the situation is far more nuanced than those three numbers suggest, and the atrocious yields are even after the chip has been downclocked and defective units fused off. To make matters even worse, the problems that caused these low yields are likely unfixable without a complete re-layout. Lets look at these problems one at a time.Number one on Nvidia's hit list is yields. If you recall, we said that the yield on the first hot lot of Fermis that came back from TSMC was 7 good chips out of a total of 416 candidates, or a yield of less than 2 percent.
