Nvidia seems to be so far ahead of the curve that they are experiencing problems that are unique in the industry. In their recent year end financial conference call, there was enough said to draw some very grim conclusions.
Today’s conference call was a near complete validation of all the things SemiAccurate has been saying about Nvidia. Remember when we asked if Nvidia could supply Apple? Anyone notice the part about dumping early 28nm capacity, and the disappearance of 28nm Fermi shrinks? Remember how 28nm was not an issue for Nvidia, even if their product roadmap slips said otherwise. How well does this mesh with the quotes from Jen-Hsun himself on the topic?
The financial update from Nvidia a few weeks ago mentioned two factors for the lowering of guidance, the shortage of hard drives, and the pre-Tegra3 drop in Tegra2 sales. The latter is to be expected, especially in light of the awful reception Tegra3 is getting from OEMs. On the other hand, the hard drive shortage that should be affecting AMD GPUs as well as Nvidia GPUs seems to only have hit one company. Funny that.
Lets look at some of the problems that surfaced during the call in more detail, as well as a few other gems that leaked out. On 28nm, the first hint at problems came out a few weeks ago when mysterious analyst reports came out of nowhere saying that many companies were having problems with yields on 28nm. This raised red flags for two reasons, first was the single source in the financial community, something that positively screams ‘planted leak’. Next is the fact that SemiAccurate talks to several companies making 28nm chips at TSMC regularly, and yield is the first topic that comes up.
There was a major change/tweak to the 28nm process last summer. Before that, companies openly grumbled about the process, and could not be described as happy campers. With the wafers that came out last summer, the frowns turned upside down. The mood wasn’t exactly joyous, but things were upbeat and yields were said to be ‘as expected’ or even good by multiple companies. Until the recent ‘revelation’ about yields, that was how it remained.
As usual, SemiAccurate immediately checked out the 28nm yield claims, and found absolutely no change. If anything, yields were better than the last time we checked on them, and no one queried could explain the reports. TSMC has publicly stated that yields are better than expected, and AMD has publicly said similar things about their 28nm experiences. For the record, Nvidia was not one of the companies queried. Ironically, that same report came out five days before Nvidia lowered Q4/2012 guidance citing 28nm yields as one of the causes. While not definitive in any way, the timing sure looks a bit odd.
During today’s call, Nvidia once again blamed TSMC for their yield problems. As we have said, Nvidia seems to be alone in having these problems. Again. The yield problems, combined with the gem that the 40nm per-chip pricing had expired or did not carry over to 28nm, dropped Nvidia’s numbers quite a bit. The view at SemiAccurate is the same as it was before, Nvidia does not have a process, 28nm or 40nm, problem, it has a management problem. Instead of addressing that side of things, once again, the company blames external forces that no one else seems to be running into.
This brings up the question of why if yields are lower than hoped for, and wafer start capacity is desperately short, did Nvidia dump all their early 28nm capacity? Related to that, what happened to the 28nm Fermi shrinks that were on the roadmap? This is pertinent because on the call, the possibility of substituting 40nm chips for 28nm shortfalls was brought up.
That answers our question about Nvidia being able to supply Apple, it looks like they will even if they can’t get enough 28nm wafers. Logic dictates that others won’t be so lucky, and that brings up some rather nasty political ramifications. If AMD can supply 28nm GPUs,one has to wonder about the design wins Nvidia is counting on for Ivy Bridge. Where the chips end up lying is unknown, but low yields and wafer shortages to satisfy high demands are not a recipe for keeping customers.
Two things will solve this problem, TSMC ‘fixing’ their ‘yield problems’ and more wafer supply. Nvidia’s 40nm woes were twofold. Early on, TSMC did have a contamination problem that affected all customers, but once that was solved, Nvidia could never get GF100/110 yields to acceptable levels. Although the company denies it, every industry insider queried by SemiAccurate pointed to design problems that were never rectified. 28nm seems to be a rerun, blaming design problems that are potentially not economically viable to fix on external forces. How much can be done and how quickly is for the moment, unanswerable.
From there, wafer supply comes in to the picture. As AMD’s Llano proved, low yields can eventually be papered over with more wafer starts. If you essentially have a fab to yourself, and wafer starts are gated by tools being rolled on to the floor and calibrated, this is a solvable problem. If you are fighting with many other players for a limited amount of capacity, it becomes an economic problem. He who is willing to pay more gets more. We will leave out the intangible cost of publicly blaming others for your own faults, it is likely unquantifiable.
Once again, Nvidia has a serious problem on its hands. It was allocated capacity and then gave it up mysteriously. How much capacity and for how long is not publicly known though, and could very well be a non-issue going forward. At the time, there were few customers that needed the early starts, a normal situation early in a process ramp. By the spring, Qualcomm is going to be ramping S4/Krait chips on 28nm, and those likely have a very high ASP per mm^2 of silicon. They will also need a lot of wafer starts too, and will be vying for the same fixed capacity.
AMD is also in the game, and the decision about where to make Kaveri and Kabini is said to be made late in this quarter. If you don’t think that will influence wafer supply right about now, please write the author for the full portfolio of for sale bridges in the vast SemiAccurate real estate holdings, the good ones are going fast.
The worst part is Apple, you know, the ones that Nvidia is likely going to burn a few bridges not owned by SemiAccurate to supply GPUs to. Apple is strongly rumored to be ramping A6 production at TSMC right about now. If true, Apple is a new account for TSMC, and they are notoriously finicky about their supply chain. On top of this, the fruit themed widget purveyor is legendary for it’s ironclad contracts that are said to guarantee preferential supply. TSMC is unlikely to step on Apple’s toes for anyone, Nvidia notwithstanding.
Add it all up, and you come to one conclusion, don’t expect 28nm wafer supply to get better for Nvidia any time soon. Same for the 28nm ‘yield problems’, design issues take months to correct if it is possible. 28nm is shaping up to be a mess for Nvidia.
The last gem thrown out on today’s conference call is an absolute shocker, Nvidia is more or less permanently out at Samsung. We say shocker not because it is a surprise to anyone following the mobile market, but because Nvidia actually admitted it in public. Samsung has had a lot of teething problems with it’s Exynos chip line, but those are very likely a thing of the past now. The company has been quite open about its plan to use in-house CPUs/SoCs preferentially, and now has the parts to do it with. For the short and mid-term, this door is closed to any volume for Tegra chips.
If you recall the Q3 financial conference call, Jen-Hsun was quite open about going after the next Kindle Fire design. SemiAccurate did some checking around that and found two pertinent details. First is the 10″ Fire prototypes floating around Taiwan late last year did not have Tegra in them. The roadmaps have definitely changed since then, so Tegra could be in the rumored 8.9″ variant. We suspect this is more of a change to the roadmap of the rumored supplier rather than actions by either Amazon or Nvidia.
On top of this is the other tidbit. A few months ago Nvidia sales people were said to be ordered to win the next Fire design at any cost. This is curious in light of Nvidia IR telling any analyst who didn’t run away fast enough how unpalatable the Fire was. The purported reason is that TI was selling Amazon OMAP chips for the Fire at $7-8, and at those prices, Nvidia was not interested in the design. TI has strongly denied these numbers, but the important point is not the actual sales numbers, it is cost.
It is SemiAccurate’s opinion that Nvidia does not have a product that can compete with TI’s OMAP in this space for a number of reasons. Tegra 3 has a bloated die and is made on an expensive process, Tegra2 has some power use figures that make it unpalatable for the Fire’s target market without massive BoM bloat, and both chips badly lag TI and Qualcomm on I/O. If Nvidia can get a win on the next Fire, it will likely come at a high cost.
That leaves the Tegra line with a very big problem, the number one and number two phone makers, Apple and Samsung not necessarily respectively, are closed to them. The number one and number three tablet maker, Apple and Samsung, this time respectively, are also closed to Nvidia. The number two tablet maker, Amazon, if technically open, is likely to have painfully low margins. The remaining tablet sales can be considered rounding error, but only if you are feeling charitable.
Phones are the remaining potential bright spot. The non-Apple, non-Samsung smartphone market is around 50% of unit sales and growing like mad. Unfortunately, Nvidia is facing a 2012 market where A15 based CPUs will dominate the high (read profitable) end, and Nvidia doesn’t have an A15 based part coming soon. On the low end, the same problems that make the Fire unpalatable apply, high costs, high power use, and bad design choices. Nvidia does not look to have any advantage to speak of in the phone market.
To make matters yet worse, Qualcomm and TI both have A15 based chips coming soon. On top of that, Intel is making a renewed effort to enter phones with their Atom line, and several others are unveiling very competitive SoCs. The once wide open phone market is rapidly becoming overcrowded, and most players have parts much better suited to it. Throw in one major player that has yet to announce their part, and things look ripe for a shakeout.
Getting back to Nvidia, the conference call that just ended was quite grim. According to the words spoken, it is a perfect storm of uncontrollable external forces. If you listen to how those words were spoken, you might notice some voices that normally don’t waver did, and volumes went way down when certain topics were discussed. This could be coincidence, but yield problems that only seem to affect one company and lopsided hard drive shortage problems say otherwise. Don’t look for things to get better from here.S|A
Editor’s note: A quick conversation with Charlie yielded this tidbit about Kepler. His previous article about Kepler did not include yield problems, which at the time of publication were not evident whatsoever. At the time he researched that story he says that everyone, including public statements by Nvidia, said there were no yield problems on TSMC 28nm. I asked his opinion on Kepler still being a winner and he said “It all comes down to execution”. Yield problems change the equation and an update will be forthcoming if Charlie gets solid answers to some questions he asked in return.
Latest posts by Charlie Demerjian (see all)
- Thing go bump(gate) in the night for Nvidia’s GP100 Pascal GPU - May 3, 2016
- Cavium’s Octeon TX blends compute and packet moving - May 2, 2016
- Qualcomm releases Zeroth API to developers - May 2, 2016
- Another detail about Qualcomm server SoCs revealed - Apr 27, 2016
- AMD finally really honestly launches the dual Fury - Apr 26, 2016