Custom is In the Eye of the Beholder
Marvell, Broadcom, Agentic AI and more
Happy New Year to everyone. We will be at CES next week, drop us a line if you would like to meet up.
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
We closed out last year spending some time looking at Marvell. They held an analyst day in late December in which they redefined themselves as a data center infrastructure company. Marvell sits in an interesting spot. They are highly capable, with some very good products. They have built good customer relationships with the hyperscalers. Seen in that light their product portfolio has a coherent order to it. But deep down we wonder what their long term position in the industry will be.
The Street has begun to grasp that the AI market is largely divided between Nvidia products and those of the hyperscalers. Both Broadcom and Marvell have positioned themselves to support the hyperscaler side of that equation. But there is still a lot of confusion as to what the “custom silicon” business looks like. This business has evolved considerably over the years and we can see the divergence in value provided by those companies and the more commodity level services offered by many others. Ultimately it comes down to what IP the various providers have on offer.
We think this could be a good year for analog semis. The industrial cycle seems to finally be bottoming out. There should be improved aerospace spending and the auto segment has some meaningful structural changes underway that favors the big analog companies. On the other hand, all these companies face a big challenge as to how to participate in AI. None of the big companies have much in the way of digital expertise and so now have to decide if they want to build, partner or buy their way in to the segment.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
TSMC employees have many more babies than other people in Taiwan. The developed East Asian countries have big demographic problems, people there are just having fewer babies. The fact that employees of one company have identifiably different behavior highlights that these demographic issues are rooted in economics and policy. Put another way, the people of Taiwan are subsidizing TSMC in some very hard to detect ways. Keep this in mind the next time someone talks about comparative productivity levels of fab employees.
The first RISC V laptops are coming to market this year. To their credit, the company building them is explicitly positioning it as a test bed product, that seems to be their core focus. The fact that this article barely mentions software is a big clue as to the broader commercial prospects such devices.
Leading Apple supply chain analyst Ming Chi Kuo says Apple will move away from the unified memory architecture it unveiled last year for the next version of its laptop CPUs and GPUs. This approach had considerable appeal, and the reasons he attributes for them eschewing it now appear to be largely around their need to move to chiplets in the M5. As such, we think if this is true, it is likely a temporary step.
Sandia Labs talks up immersion cooling. Most people we speak to about cooling data centers say that immersion cooling is likely inevitable, AI systems are just getting too hot. But almost everyone is dreading the prospect. We think the industry will push liquid cooling as far as it can go, and it still has a few more cards to play. But given the power specs we have been hearing for future Nvidia systems later this decade, the industry is just going to have to figure out how to get immersion systems to work.
Nikon unveiled a lens that can take both wide angle and telephoto images at once. That would seem to defy the laws of lens physics. Fortunately, they will be at CES and we can see for ourselves.
This company sells a box to Airbnb hosts that claims to detect how many people are in a house. The idea being that Airbnb hosts do not want parties in their properties, and the “Party Squasher” will nip those parties in the bud. The more we read about it the more we came to see it is sitting on just this side of legal, but definitely creepy.
Digitimes talks up Nvidia’s push towards Ethernet for its future AI training systems. The fight over AI networking is going to be highly contentious this year.
Networking and Wireless
Every generation of wireless standards (the ‘G’s’) Qualcomm and others try to broaden the market by launching products for ‘unlicensed spectrum’, frequency that is not controlled by the operators. 5G is no different, with Qualcomm announcing its support for 5G NR-U. They have done some smart work and combined multiple past attempts for a unified approach. Full points for engineering, but we have yet to hear of anyone looking to make actual products for this.
As ubiquitous as Bluetooth has become, the Ant+ standard still exists. It is primarily used for fitness devices, and Garmin remains the de facto owner of the standard. Ant had certain advantages over Bluetooth around low power operation, even in comparison to Bluetooth LE. But its days appear to be finally coming to a close, interesting case study in how standards are born, grow and then end.
Wi-Fi networks talk to a lot of other servers than the ones users direct them to.
Software and the Cloud
Apple Intelligence’s memory footprint is growing sharply. The software suite is evolving and the company is asking users to dedicate ever more amounts of capacity for it. Recall that Apple Intelligence is only available on a subset of iPhones that come with sufficient DRAM, essentially just the top models. This mission creep demonstrates that even a company like Apple, who knows better than anyone how to turn technology into products is still struggling with that process in the AI era.
This author laments that the role of Developer Evangelist is dying. His argument is that growth imperatives within venture-backed SaaS companies has kneecapped what was once seen as a critical role. Anecdotally, we have seen this in action. We link to it here because we write a lot about the importance of software to semis companies, and it worries us that if software companies are losing this function, what hope is there for semis companies to pick it up just when they need it most.
“Agentic AI” is going to be a big topic this year. This company walks through how they built their agentic system, with specialized LLMs querying other specialized LLMs. We think wider adoption of these is still further out, but it is interesting to see the design space that this approach opens up.
Large Language Models are premised on the computational idea of “attention”. This author argues that attention can be analyzed and managed in much the same way as we think about electromagnetic fields. The RF portion of our brain finds this deeply satisfying, even if the math hurts the rest of our brain.
Climate and Science
China has reduced sulphur emissions by two thirds. Anyone who has visited China recently knows the extent to which the chronic pollution of even a few years ago is much improved.
Diversions
If you are shopping for a late gift, and could not get your hands on the ASML EUV Lego set, you can now get a Lego set of Ernest Shackleton’s HMS Endurance.
Giant, 1,000 year old cities buried underneath the jungles of the Amazon.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.