D2D Highlights April 10, 2023
Who is really the commodity player here?
Highlights from our blog
For years, people making “serious” semis have neglected sensors as almost an after-thought, a sleepy corner of commodity parts. That is changing. Advances in sensor manufacturing, borrowed from advances elsewhere in the industry are leading to big gains in sensor performance. At the same time, improvements in software (aka AI) make all that sensor data much more useful. Finally, new business models are opening up around services and solutions, which means sensors may finally be able to capture more value. There are now a dozen companies designing CPUs, but we know only one company capable of making ultra-high precision position and current sensors at affordable prices. Remind us, which one is the commodity?
Amazon opened up its Sidewalk network to third party developers. Sidewalk is probably the largest example yet of an ad hoc network, a topic which we find endlessly fascinating. These could be incredibly useful for really expanding the definition of the Internet of Things to trillions of objects. That being said, no one has yet figured out how to make ad hoc networks commercially viable.
Other Items of Note
Semiconductors and Deep Tech
Google released a paper and a lot of data on the latest version of their TPU v.4 AI accelerator. The system boasts impressive performance, but also shows how much of an arms race this is for the whole industry. They benchmark against Nvidia’s A100, but H100 is on its way. Also, very important to note that much of Google’s performance is dependent on their proprietary-ish networking stack. (Go read Dylan Patel’s analysis of that networking stack.)
Ampere released a line of demo units. For only $2,000, anyone can kick the tires on compute of the future. This looks like a toy, but it is a really smart move by the company. The price point is low enough that ‘anyone’ can try it out, run their software tests on it, and when they realize its capabilities go out and purchase a few hundred million dollars of it for their production cloud. Seriously, this is a smart marketing move.
A primer on Quantum Computing. Everybody loves the idea of quantum computing, it is incredibly fascinating. The industry has made some great strides technically, now we just need to find paying customers.
Intel made a 360-degree video of the insides of their fab. Best viewed on a phone with an accelerometer. Say what you will about Intel, they still know how to make great marketing videos.
Chips with Lasers on their heads? Is that too much to ask for? Four ways to integrate photonics and semis.
Asianometry just launched a great series on semis packaging.
A fascinating story of a debate in the race to build better software for designing chips. This does not portray the industry in a fantastic light. Spoiler alert, the protagonists end up quitting their jobs and moving to LLM company Anthropic, more talent lost to the software industry.
A look at how Mexico’s auto industry is re-tooling for the age of Electric Vehicles. If the US wants to onshore auto manufacturing, Mexico could play a major role.
A good debate on the future of EDA tools for RISC V.
Restoring the US Vacuum Tube industry. Why not?
Wireless and Networking
Reuters did a deep dive (pun intended) on the growing friction between China and the US on the subject of undersea fiber links. There is a battle underway for the infrastructure of the future, and it is a lot more complicated than just China vs. USA.
SpaceX’s Starlink is starting to ration capacity. We are fans of Starlink, it is quite a technical achievement. But a few years ago, everyone seemed to think it was the solution to Everything. It turns out that networks run really fast when there only a few users, and bandwidth is finite.
Software and the Cloud
On the last episode of The Circuit, Ben Bajarin and I talked about the struggles auto makers will have building software for their vehicles. Ars Technica just took a look at Volkswagen’s attempt to do exactly that, and not surprisingly, they are not having an easy time with it.
We recently asked a friend with deep AI roots for a primer on LLMs and other recent advances in machine learning. A day later Stanford released a 300+ page report on exactly that subject. And here is a quick Twitter thread summary. More on this topic soon.
Like everyone else, we are running to keep up with all the advances in Large Language Models (LLMs). For semis, a big part of the story will be how do LLMs run on edge devices like phones and laptops. There is already a lot of work being done to make Edge AI faster.
Other Items we Found Interesting
A cybersecurity curriculum from the National Security Agency.
Nasa has developed a new, more efficient rocket engine. Not quite the Epstein Drive from the Expanse, but still fun to see.