Rethink Technology business briefs for September 21, 2017.
A Model S P100D wins the quarter mile against . . . just about everyone
Source: Motor Trend
Although I can’t recommend Tesla (TSLA) as an investment at the moment, I’m still a shameless fan of the company and its products, and happy to report good news when it’s available. Motor Trend recently held what it called the “world’s greatest drag race” on the landing strip at Vandenberg Air Force Base in California. And a Tesla Model S P100D with Ludicrous mode won the race.
The Model S’s competitors included some potent and fairly exotic machines such as the Ferrari 488 GTB, Mercedes AMG GT R, Aston Martin DB 11, and McLaren 570GT. The P100D often wins such drag races, where its electric motor torque pushes it to the quarter mile finish first.
But at longer distances, the Tesla usually falls behind, since it doesn’t have the top speed of the exotics. But it was still fun to see the competition trail in the wake of the mighty P100D. And races such as this demonstrate the future of the performance sedan: electric, all-wheel drive. Whatever may befall Tesla and Musk, even its detractors have to admit that Tesla has shown us the future.
Tesla’s reported custom AI chip: that’s what Keller does
Yesterday, CNBC reported that Tesla is working with Advanced Micro Devices (AMD) on a custom AI chip. Jim Keller is reported to be leading a team of “more than 50 employees” for the project. Tesla reportedly has received samples of the new chips and is now testing them.
GlobalFoundries, the fabricator that does most of AMD’s work, reportedly made the parts. CNBC says that GloFo’s CEO Sanjay Jha mentioned the work being done for Tesla, although this was seemingly denied later on.
Reuters subsequently reported an email statement from the company:
Tesla has not committed to working with us on any autonomous driving technology or product.
Of course, there wouldn’t be a commitment at this point, but this statement isn’t a denial that some work has been done. Ever since Jim Keller was hired away from AMD in January 2016, rumors have swirled around him that he was designing a custom AI chip to power Tesla’s self-driving cars. Jim is a microprocessor architect. Designing chips is what he does. It’s probably not plausible to assume that Tesla hired him for any other reason.
However, designing microprocessors from scratch is a huge, billion-dollar undertaking, something that non-technical business writers may not appreciate. My take on the rumored Tesla chip is that a collaboration with AMD was always the plan. Even with Keller and his staff, many of whom came over from AMD, Tesla wouldn’t have the resources to go it alone.
So I think it’s likely that Tesla hired AMD to design a “semi-custom” chip along the lines of the console chips. This is the somewhat mysterious “third semi-custom design win” often referred to in AMD conference calls.
Keller’s staff are overseeing the design effort and providing design input. One of those who Keller brought over to Tesla, David Glasco, is listed on his LinkedIn resume as System Architecture Lead at Tesla. But he never left the Austin, Texas, area where AMD is located.
As to the composition of the chip, many have been assuming that it contains custom ARM CPU cores. I actually think this is unlikely for a number of reasons. AMD doesn’t really have the skills to design a true custom ARM core.
But the most important consideration driving this development process for Tesla is the desire to find a lower cost solution than what NVIDIA (NVDA) has to offer. Capitalizing on the development of Ryzen and Vega seems like a good way to do that.
So my take is that the chip is probably a combination of one or more Ryzen “Zeppelin” slices, combined via the Infinity Fabric with a Vega GPU and possibly a custom ASIC for hardware tensor processing. This would take advantage of AMD’s development of Infinity Fabric for EPYC and Threadripper, and make the resultant device relatively low cost to fabricate.
Trip Chowdhry begs to differ
Barron’s reports that Trip Chowdhry of Global Equities regards the CNBC report as “100% false.” In a research note, he writes:
Comprehensive view is that TSLA and NVDA have currently a 5-year contract in place, which was renewed just recently. AMD is not a Player. AMD is DoD (Dead-on-Departure) in DML (Deep Machine Learning) Workloads. Google threw AMD out of its GPU Training Cluster, as AMD had extremely poor performance on GOOGL TensorFlow framework. Later, as a courtesy to AMD, GOOGL redeployed the AMD GPU’s for only VDI (Virtual Desktop Infrastructure). So far we have attended no less than 60 DML (Deep Machine Learning) conferences…we have not seen even a single benchmark of any DML Framework running on AMD GPU’s for production workloads. Investors optimism is completely misplaced that AMD will become a significant DML player. There is only going to be one GPU player NVDA, just like there is only one CPU Player INTC ….rest all the players will be in the others category, which will be about 10% of the market.
Insofar as AMD’s disadvantage in AI on its GPUs is concerned, I don’t doubt that Chowdhry is correct. But I doubt that would stop Keller from pursuing this project. And even though there might be a contractual commitment to NVIDIA for some period of time, projects like this require a significant amount of time. Tesla may simply be looking forward to the next generation of devices post-NVIDIA.
The area where AMD is weakest compared to NVIDIA, in software support for machine learning, is precisely why this effort may be on a multi-year development track. Getting to the point of having a hardware platform is just the start. Now the real work begins to develop the software.
While I disagree with Chowdhry on the reality of the effort, that doesn’t mean I think it’s a good idea. I’ve written previously that Tesla’s autonomous vehicle effort appears to be in disarray. Now the development of a separate hardware platform and the concomitant software effort seems like grasping at straws.
Tesla hasn’t been able to make much progress on the current NVIDIA derived platform, which is half of a Drive PX 2, and I believe, inadequate to support full self-driving capability. So Tesla has decided to go off in a completely different direction. I believe that Tesla would have been better served devoting the money and resources from the AMD effort to solving the problems it has with the current system, whatever those are.
That just wasn’t going to happen once Keller landed on the scene.
Nvidia is part of the Rethink Technology Portfolio and is a recommended buy.
Disclosure: I am/we are long NVDA.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.