Nvidia took a decidedly different tone at the launch of Tegra 4, and the muted tone shows exactly how much trouble their mobile line is in. If you look back at history and throw in a few more recent data points, the situation looks dire.
At CES this year, Nvidia didn’t announce the tens of design wins that Tegra 2 theoretically announced, the few dozen that Tegra 3 triumphantly boasted, or the magical qualities of the chip that would power it to market dominance. In fact, the number of design wins announced at the launch, not counting internal reference designs, was a rather sobering zero. Instead of partners lining up to build things on it, we got in-house products with a rather tenuous future.
How did Nvidia go from being the industry savior to peddling odd game consoles? How did their graphics leadership evaporate? How did their unquestioned phone advantages vaporize? How did they go from the top of the world last year to the bottom this? That is easy, the company never had any of what they boasted, and what little was actually promised to OEMs fell short, so short of the desired goals that there were repercussions. As SemiAccurate has been chronicling for years now, mobile device OEMs can be lied to once, sometimes twice, but the third time is not going to happen. Sometime after round one but before round three, the sales team does not get calls answered.
Nvidia has long been one to boast about their own advantages, market dominance, and inherent victories, and most press outlets are rewarded well to repeat those claims while never questioning predictions that don’t materialize. How many of the dozens of Tegra 2 and Tegra 3 design wins ended up as devices on the market? Why again is their first A15 based SoC a 2013 product when everyone else has had A15 devices on sale for quarters? Why can’t Nvidia’s SoCs do OpenGL when ARM/Mali and Imagination can? Where are the Tegra 3 phones that were promised to hit the market in force? Where are the tens of Tegra 4 design wins Nvidia normally rattles off to every analyst that can’t run away fast enough? What happened to the bravado at CES?
Behind the scenes, a remarkably different story from the public bravado was playing out around Tegra, and it was not one with a happy ending. The pattern started with Tegra, no numerical suffix, the chip was Nvidia’s first foray into the phone and tablet space. During Hot Chips 20 a few years ago, Nvidia put on a talk about Tegra and emphasized it’s HD video capabilities. 1080p was touted as the killer app, no other chip on the market could play video well at that resolution. They showed movies and games, but the displayed picture was not good quality. This was blamed on the projector, a problem that plagues public demos worldwide.
Checking the specs in the presentation, specifically page 5, it appeared that the device could not actually support the promised 1080p resolution on an internal display. When asked if the presenter meant 1080p video on an external display, they said that Tegra could not support that resolution on it’s external output either. Further pressed by SemiAccurate about how the chip could claim to support 1080p video if it could not show it on an internal LCD or external screen, he admitted that in Tegra, 1080p was, “just an internal decoding format”. Imagine if you were an OEM that had committed to Tegra devices based on the promise of 1080p video, and were told that it was “an internal decoding format”. Some still question why this part had little uptake.
Tegra 2 is another piece of the pattern. The dozens of design wins promised at CES? Poof. Why? Sources at the time said that the silicon delivered was roughly 25% over the promised power budget, an unforgivable ‘misstatement’ in a space that will throw an SoC out the door for 10 milliwatts too many. To make matters worse, that 25% was at a far lower clock than originally promised, the promised clocks would have unworkable power draws for a mobile device. The cherry on top was that the two promised killer features that no other device had, interrupt driven USB and one we were asked not to reveal, both simply did not work. Luckily, at least according to what Anand was told in the above link, it was all due to OS choice, not a 25% power miss and a large clock miss and both ‘killer app’ features simply not functioning on a device that missed it’s promised ship date by a wide margin. OS choices were the publicly stated cause and people didn’t question it.
Step up to the next generation, Tegra 3, the one where Nvidia finally showed what it could do. It had more design wins than Tegra 2, at least according to CEO Jen-Hsun Huang in a Dow Jones piece, so it would dominate the market, right? SemiAccurate said it was a badly designed dog that once again missed performance goals by wide margins. Their CEO was telling the financial and technical press how wonderful it was, Mike Rayfield was claiming claiming 30 design wins, twice the 15 that Tegra 2 had. Tegra 3 was a winner, and had tons more un-nameable design wins, and many features that would put the competition to shame, just you wait and see. The OEMs waited. We waited.
The flagship Tegra 3 phone by HTC had a Qualcomm SoC in it when it was released in the US, the rest of the design wins never broke cover in a meaningful way, dying quietly in the dark with no explanation. Did you ever see a Tegra 3 device in the wild, they were rarer than Zunes at MacWorld. While some probably did make it to market, they were so few and far between that you could be forgiven for not noticing.
Anyone testing power numbers in actual devices had a very hard time getting anywhere close to the power numbers claimed by Nvidia, that was a clue. It was dismissed by the company as having been a specific device or testing method, every one of them that every tester tried mind you seemed to be an anomaly.
The power numbers in delivered silicon, while better than Tegra 2, fell short of the spec sheet in independent testing.
For Tegra 3, two design wins were saving graces, the Microsoft Surface RT and the Google Nexus 7. Surface RT sales were so abysmal that Microsoft would not even give out a number to analysts, and the replacement for the Nexus 7 gave the design win to Qualcomm, not Nvidia. Shades of HTC? Given how bitterly Nvidia representatives were badmouthing Microsoft to financial analysts during most of 2012, it is hard to see why they were publicly happy about the Surface.
You might see a pattern developing. Nvidia promises X for the new CPU. OEMs like what they hear because the promised specs are genuinely good and sign orders. Next those customers start on designs with those upcoming SoCs before silicon is delivered, something that may seem odd but is a necessity with modern design cycles. Work well underway, Nvidia delivers early silicon to the OEM. The promised power draw, speeds, and feature sets are a pale shadow of what the OEM had in signed contracts. Meetings ensue and assurances are given that this is just early silicon, all will be as promised after a few steppings. The OEMs grumble loudly but carry on because Nvidia would not flat out lie to them?
While SemiAccurate has not attended any of those meeting or seen any direct paperwork resulting from them, we have talked to people who were directly involved, and they all described events in a very similar manner. They also go on to say that once final silicon was delivered, it did improve a bit, but did not resemble the initial promised parts at all. By the time of the first panicked meetings, designs are so far advanced that it is too late to change gears. If an OEM did not have a full plan B for the device being developed in parallel using another SoC, a costly and time consuming endeavor, there is no choice but to carry on. Dumping Tegra at that point would lead to several quarters of delay for the product, or worse yet nothing at all to bring to market that season. By the time final silicon is delivered, the device is quite final and only debugging remains. If Plan B wasn’t similarly fully baked, they just have to live with Tegra as is.
The dozens of design wins Nvidia promised for Tegra 1, 2, and 3 that never materialized are indicative of just how bad the delivered silicon actually was. This isn’t some critic taking potshots at the company, it is a multi-year track record of unprompted statements by Nvidia PR, management, and engineers that simply did not happen. Tegra, Tegra 2, and Tegra 3 all had specific numbers of ‘won’ designs promised, and all had a minute fraction of those materialize. It happened again and again, the record is quite public in trade press, financial outlets, and even videos on Nvida’s own site. SemiAccurate wrote about many of these and Nvidia was quick to counter through tame or paid media outlets. That said, a few stories did get out, remember Boxee, the marquee design win for Tegra 2?
With such a long and consistent track record, there are two conclusions that could be drawn. The first is that Nvidia spokespeople, management, and engineers are flat out lying to press, analysts, and customers. Second is that the silicon delivered is as bad as SemiAccurate says that it is, and the bugs and spec misses are quite real. Looking at the Boxee story on Anandtech and comparing the clock speeds of Tegra 3 to what was promised a year earlier, there should be little doubt that the delivered silicon falls woefully short. Spec sheets and quarterly financial calls are pretty hard to dispute after the fact.
Readers can make up their own minds about Nvidia management’s truthfulness, but you might want to go back and look at what was promised for Project Denver at CES 2011. Nvidia CEO Jen-Hsun Huang said that it was three and a half years through a five year design cycle in March 2011. SemiAccurate said there was no chance of that and he knew it.
It is now the end of the time frame Nvidia management gave analysts for Project Denver’s release, individually and off the record though it was, it was still a hard date. We told analysts at the time that they were being lied to on purpose, but what was our inside word against direct CEO and IR statements? No points if you said correct. We will happily eat these words if Project Denver silicon ships in the next 6 weeks though, there is still time to be proven wrong, technically anyway.
As SemiAccurate has been saying for years that it is easy to burn OEMs in the phone and tablet space, they are a trusting lot. If you have even a bit of credibility and offer them a plausible killer chip, they will jump. Unlike desktop oriented silicon, phones have a rigorous testing and delivery schedule, you had damn well better deliver what you promised. If you don’t, the OEMs and carriers will find out during testing, and the FCC has a notably droll sense of humor about problems that make it as far as their labs. Making a device around an SoC is a long and expensive process that ties up huge amounts of engineering resources. If an OEM commits on a promise and it doesn’t turn out, a lot of money is wasted. Worse yet, a company has to go to market without a competitive product, if they have a product at all. If you think the FCC’s sense of humor is droll, OEMs and carriers make them look like stand up comedians when things like this happen, and they have long memories.
In light of this, Nvidia burned just about all their potential Tegra customers, some quite badly. Tegra 2 was the first device that was really aimed at phones, and there were a lot of customers lined up. Tegra 2 failed to deliver in a painful way on power, performance, performance per clock, performance per watt, and promised feature sets. It just didn’t work right. All those “dozens” of design wins that went away were not absentmindedness on the OEM side, there was a reason for it. Tegra 3 had a much lower number of OEM and carrier design wins, PR promises aside, and Nvidia woefully underdelivered once again. You can burn them once (Tegra), you can burn them twice (Tegra 2), and you can even burn a few a third time (Tegra 3), but most won’t pick up the phone at that point.
Tegra 4 had zero non-reference design wins announced at launch. This wasn’t a fluke. This wasn’t a change of PR philosophy. It wasn’t conservatism to win back the stock market oriented crowd, sources say that the number of design wins was minuscule. There was a Visio tablet shown off and a rumored LG win, but adjectives like dozens and tens were notably absent for cause. At this point, Nvidia has burned all the players at least once, and the results are quite clear to anyone looking on, Tegra 4 has few if any real customers.
This may come as a surprise to onlookers, but to Nvidia management, the problem has been known for quite a while. The CES keynote was good evidence of the lead times they had on the chip’s uptake, you can tell by the devices shown. How many non-Nvidia Tegra 4 reference designs were used in Nvidia’s keynote? Instead of killer phones or tablets from volume OEMs, you got an odd handheld game controller called Project Shield. It is more evidence of Nvidia’s increasing desperation to stay relevant in the game industry after losing all three next-generation consoles to AMD, but even that took months and months to finalize. It wasn’t a snap decision, but it gives you a nice window on what management knew when in regards to Tegra 4 uptake.
All of this news is history, and that usually is seen with 20/20 vision unless your main advertiser would disapprove. Unfortunately, two recent rumors that after extensive research, SemiAccurate believes to be true, paint a much more dire picture for Nvidia’s entire Tegra line. To be blunt, the chances of any significant device wins are essentially zero, and Nvidia is actively taking steps both stem the bleeding and keep volume up to economically viable levels. Because both are short term desperation plays that will have serious long term consequences, we have no doubt that management would only attempt them if there was absolutely no other choice. They see the end approaching, and are trying everything to avoid that fate.
Note: The following is for professional and student level subscribers.
Disclosures: Charlie Demerjian and Stone Arch Networking Services, Inc. have no consulting relationships, investment relationships, or hold any investment positions with any of the companies mentioned in this report.
Latest posts by Charlie Demerjian (see all)
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016
- Intel’s Xpoint is pretty much broken - Sep 12, 2016
- ARM adds 2048-bit vectors to v8A with SVE - Sep 7, 2016
- AMD releases Bristol Ridge 7th Generation APU - Sep 5, 2016