D2D Apple's modem, RISC V, lots of CPUs, and AI
Become a Paid Member of the D2D Community
Just a reminder that we will soon be shifting our newsletters to 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.
Check our podcast - this week we welcome our first guest Calista Redmond, CEO of RISC V International.
Highlights from our Blog
The market assumes that Apple will replace the Qualcomm modem in the baseband as soon as its internal modem is ready. But Apple has a poor track record with networking chips. Case in point, its internal Wi-Fi/Bluetooth chip has only seen the light of day in the Apple Watch. We give Apple Silicon the benefit of the doubt, but success is not guaranteed.
Nvidia is obviously doing very well with its chips for AI, but Qualcomm now has a fairly solid AI stack. Unfortunately for them, they will have a harder time monetizing those capabilities as “AI” is likely to end up as a feature in their existing product rather than a new product category. As the competition heats up for AI at the edge, this should be an important lesson for other, aspiring entrants to this very large market.
In the past month we have seen dozens of pitches for companies highlighting their CPU and NPU roadmaps. They are getting very hard to tell apart. Every one seems to have the same slides showing how their solution stacks up against the competitors. Increasingly, technical details - feeds and speeds -matter less than business models and go-to-market plans.
We often get asked what would be our “ideal semis start-up" would look like. It would need to blend solid hardware with a '“software first” mentality, and through this build a company with some form of true competitive advantage.
Inventory levels are still high. Margins are under pressure from inflation. And consumer demand looks highly questionable. So of course semis stocks are up 40% this year.
Noteworthy Items
Semis
Several leading RISC V supporters founded an organization, called Rise, to promote software for the platform. Most noteworthy are Intel and Google’s leading roles, as well as the inclusion of Mediatek, Samsung and Nvidia on the founding board. On the latest episode of the Circuit Ben Bajarin and I talked with the head of the RISC V International governing body. We see efforts like Rise as further demonstration that large companies are taking RISC V very seriously.
Arm announced the latest updates to its processor roadmaps. Keeping track of Arm’s product nomenclature is always challenging, but our big takeaway from this event is this is largely more of the same - incremental improvements to its offerings. Reading between the lines, and listening to industry chatter, our sense is that company is not taking any chances ahead of its IPO. Many would have liked to see a more aggressive push from the company across many fronts, but they are likely constrained by Softbank’s goals of keeping expenses down to help valuation.
One company’s progress on porting their application to run on RISC V. Put simply, it runs but will take some work to optimize and get it competitive with x86 run-times.
An interesting look at the changes to Intel’s chip topology. When (if) Intel finally fixes its manufacturing process it is going to have to start competing on its design and architecture capabilities again.
The Next Platform thinks we are seeing the bottom of Intel’s data center decline. Maybe they are right. Maybe not.
Hardware
Tesla has rethought the entire process of building cars, with a manufacturing system that turns traditional practices on their heads (almost literally). This has major implications for their costs, and bolsters our thesis that the shift to EV could entirely disrupt most of our assumptions about the costs of making cars. To get a historical perspective on the difficulties of designing profitable manufacturing lines, this look at the history of automotive paint is much more interesting than you would expect.
Electrical grids have become furniture, items we largely take for granted. But it took a lot of work to get them to the level of reliability we enjoy today (mostly). China shows that the transition to new energy sources works best with a major upgrade to those grids. So it is worth thinking through the history of the grids and simple-sounding inventions like Power Relays which play an important role in the grid, despite few people knowing about them or the sizable company that supplies them.
Networking and Wireless
More news for the Open RAN initiative. Nvidia is teaming up with Softbank to provide a networking solution. Meanwhile, AWS is positioning Graviton to provide those same functions. These are both bad news for Intel’s efforts here.
We are closely watching the market for private networks. DFW airport recently held a big process to select a vendor for theirs. This is a case study in how an enterprise like this can benefit from a private network, that ties together their all their networking needs, including integration with Wi-Fi. Also, a white paper from Dell on their private wireless solution. It is getting easier to build these networks.
There is a rumor making the rounds that Amazon is going to offer ‘free’ mobile phone service as a benefit of Amazon Prime. Everyone involved has denied this, but it is worth pondering. Amazon seems very active in spectrum and every few years there is some new theory as to what they ultimately plan to do in cellular, if anything. It would be fairly easy for them to enter into an MVNO agreement and make an offering like this. That being said, what makes sense for one part of Amazon would be fairly detrimental to the AWS side which is trying to offer telco cloud to companies the other side would be competing against. As Amazon grows into every corner of the economy it will face more of these strategic dilemmas.
Software, AI and the Cloud
Part of our struggle with the latest Large Language Models (LLMs) and Generative AI is that we struggle to see how consumers will actually use it. Apparently, building useful UIs on top of LLMs presents a whole other set of challenges than just getting the models to work at all. As this write says “LLMs are not products”.
Compute in general, and AI in particular, are straining what data center operators and Co-Lo facilities can offer. Many eye opening comments here, including the fact that data centers now need to provision 5 kilowatts per server cabinet.
Other Interesting News
Periodic reminder to manage your URLs. The state of Maryland forgot to renew one of its URLs and now their license plates are advertising an online casino.