Blue Horse Shoe Loves Intel
More developments around Intel, good numbers from Nvidia and Marvell, and our time at Hot Chips.
In which we explore Nvidia’s results through the lens of their share of the data center processor wallet, which has now reached 82%. We also think through some of the activity swirling around Intel as well as review Marvell’s good print and AMD’s acquisition of ZT. And much more.
Just a reminder that we operate our newsletters on a paid model. Paid subscribers will get three newsletters a month including our China Deep Tech notes. Paid subscribers will also get early access to the newsletter as well other benefits coming soon. Please subscribe and support our work.
Highlights from our Blog
Intel is in play. After last quarter’s disastrous earnings a storm circling around the company’s Board. We think there is a good chance of an activist attack on the stock. And the idea of splitting Products from Foundry has once again reared its head. The company went through a big internal debate over a year ago about this and we thought the issued had finally been buried, but it is back again. We think a split like this will end up destroying value and probably breaking both sides. But Blue Horseshoe loves Intel.
Nvidia reported good numbers, but apparently not good enough because the stock is down for the week. We warned a few months ago that there is a big difference between Nvidia the Company and Nvidia the Stock. The stock’s trading pattern is not always going to make sense. On the other hand, the Company is doing well. Nvidia now accounts for 82% share of the data center wallet. Nvidia’s growth is slowing, but it is still strong and the run they had had over the past 18 months was extraordinary.
We were getting a little worried, but Marvell reported really good numbers. The company’s core products were hit hard by last year’s inventory cycle. As a result investors were starting to overly-fixate on their custom ASIC business, and we are a deeply ambivalent about that segment. Fortunately, the core products bottomed out, and they are enjoying a great cycle for the optical products. They are not out of the woods strategically, but we will take a good quarter for now.
AMD acquired ZT Systems. Strategically, this deal makes sense. They are competing with Nvidia and Nvidia sells complete systems - chips to racks (and even to data centers). So now AMD will be able to offer complete systems. At first, we thought the $5 billion price tag was a bit rich, but after speaking to Forrest Norrod on the Circuit, it sounds like they should be able to recoup a fair amount of the deal when they sell off the manufacturing business.
By far, the highlight of last week was the uber-geeky Hot Chips conference. We enjoyed deep looks at the real complexities of building AI systems at scale. The most noteworthy presentation came from OpenAI, but not noteworthy in a way the company hoped. “There is no way to predict the timing for AGI”, is quite a statement from the company that seems built on just that. Is the air escaping the AI bubble a bit?
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
A moment of silence for AnandTech.
We field a lot of questions as to why is Nvidia’s CUDA so important to AI. The answer is that CUDA provides all sorts of tools and hooks into the silicon that let users make the most of their silicon. Here is one example of how that works as this company set out to understand why their GPUs were not running at full utilization. In this case, even those tools needed some help, and Nvidia is way ahead of everyone else in this regard. Scaling AI clusters is a lot of work.
If there is one company that can truly challenge Nvidia anywhere in AI, it is Broadcom in AI networking. Nvidia’s networking business is now on a run rate over $12 billion a year, but the market for Ethernet semis is important to Broadcom too.
CNBC’s Katie Trasov got to visit Google’s TPU lab.
Hindenburg Research, the infamous short-selling shop, has come out swinging against SuperMicro. They call out dubious accounting (and dubious accountants), quality problems and a very questionable culture. We do not know SuperMicro well, but people we know who do are not surprised by this.
Liquid cooling is becoming a hot topic, no pun intended (well, maybe a little intended). There is a lot of noise in the system about the topic, expect more.
Networking and Wireless
ISPs offer customers in-home Wi-FI. This sounds great, someone else can manage the IT headaches. But of course, the ISPs are also harvesting way too much information from that Wi-Fi router sitting right in the middle of your living room.
Software and the Cloud
A big question in AI research now is the extent to which these models scale with data. Is more data always better? Do results scale exponentially with larger models? This author argues that the ‘true’ exponential curve can actually be found in the cost of the models themselves, they get exponentially cheaper very quickly. Some time soon, there is going to be a massive game theory at work as companies play chicken to see who will spend infinity billion dollars on an Nvidia cluster and who will come second and replicate those results on a bank of Raspberry Pies and an Apple Watch.
The current generation of AI coding tools is not a panacea. It can help productivity, but developers have to devote a lot of that time saved to reviewing the AI-created code. That can cause enough headaches to lead many of them to prefer going without the AI tools.
The RAND Corporation says 80% of AI projects fail, and when has RAND ever been wrong? This report is interesting but also a Rorschach test for readers.
How to build a spreadsheet in Python.
Science and Climate
Solar panel costs continue to plummet, but costs of solar systems are actually up a bit. The cost of everything else in the system (land, regulations, labor, etc.) are starting to matter more. That being said, solar is essentially the cheapest source of energy on the planet now.
The US (and to a lesser degree the global) economy are now much less energy and carbon intensive than they were even a few years ago. This is good news, and it turns out there are a lot of policies out there that work and can bend the curve even further.
Diversions
A mechanical neural network, because why not?
Resources for robotics hobbyists.
If novelists wrote bug reviews.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.