D2D: AI =? GPUs, Hardware + Software models, and our thoughts on Intel
There is a lot happening in AI and the data center, we look at the news as well as examining where we could be wrong
Become a Paid Member of the D2D Community
Just a reminder that we are 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.
Highlights from our Blog
We have been getting asked a lot about the future of Intel, and so we put our thesis in writing. Put simply, the company faces three sequential problems: 1) Fix its manufacturing process; 2) Fix its product roadmap; 3) Find foundry customers to keep their fabs filled allowing them to keep all of the above moving. Tough, but not impossible.
Part of the inspiration for the post above was the webinar Intel held for investors to unveil their new reporting structure. This breaks out the manufacturing operations with its own P&L. The event was both encouraging - this is a necessary first step, but also reminded us that Intel’s biggest problem is fixing its own culture. For example, they claimed that according to this new accounting treatment, Intel will be the second largest foundry in the world. They achieved this by counting the other side of Intel as a customer. Like taking your sibling to the prom, this is true but may not send the message they think it does.
In all the talk of hardware companies trying to add software products we provide the cautionary tale of CalAmp which tried to do this repeatedly over the years. They finally seem to have found the right model (and team) but they lost a lot of value on the way. More broadly, we generalize this point with what we consider the key elements needed to make this possible. This includes the obvious like have the right management team, but also pay close attention to the integration of the engineering teams. There is no quick way for established companies to do this, but it looks like a lot of hardware companies have begun the decade-long process to get there.
Amidst all the excitement around Nvidia’s earnings, the industry is coming around to the realization that AI does not equal GPUs. We will end up using a lot of CPUs, accelerators and even FPGAs for AI inference.
Listen to our Podcast
Noteworthy Items
Semis and Deep Tech
AMD has launched a massive FPGA cluster for emulating chip design. Interesting on two fronts. First, a reminder that most leading semis companies design FPGA-based versions of their chips to test them ahead of tape out. Second, interesting to see AMD steal a page from Nvidia with a move into providing Cloud services.
TSMC is reportedly planning to charge $25,000 per wafer for its 2nm process (coming in 2025). We would not be surprised if this price ends up even higher.
As computer architecture gets more complex and data center systems get more heterogenous our ability to manage and secure them gets harder. These researchers explored how to think about these systems and propose a new approach. Worth linking to if for no other reason than the title “Putting out the Hardware Dumpster Fire.”
A look at AMD’s path to developing chiplets from a member of the team who designed them.
The Japanese government is nationalizing JSR which makes some important chemical components for semiconductors among other things. This is a curious move, as past Japanese government investments in companies tended to be for troubled companies (which JSR is not), essentially converting those companies into jobs programs. But now many of those other companies are re-awakening to become commercial again, yet for some reason JSR is going the other direction. Tim Culpan takes a stab at this head-scratcher, concluding that this seems to mark a new phase for Japan’s Industrial Policies, with more consolidation coming. As governments everywhere take semis more seriously, we should not be surprised if they implement unexpected policies.
Wireless and Networking
The iPhone is gaining share in Southeast Asia. This has long been a stronghold of Chinese Android phone makers. There must be some limit to Apple’s share gains, but we have no idea where that is.
Google is removing open source access to the default Android dialer and text messaging apps. Probably necessary for security reasons, but we do not think this is the direction Google should be headed.
The tower industry held a conference earlier this month. The US 5G build out is largely over, carriers are turning their focus to fiber, which is not great news for the tower companies. mmWave is not at the top of anyone’s priority list, but there is growing interest for indoor solutions. Private networks are also starting to attract more attention.
The Penske Formula E racing team got caught using an RFID scanner to read data off their competitors’ tires. And a car battery monitor was found to be tracking users’ location and sending that data to someone, somewhere. We are only just beginning to grasp the security issues of the Internet of Things.
Nokia pushes ahead with Arm-based network processing. This stands in contrast to Ericsson which is ostensibly betting on Intel’s approach… for now.
Oh look, Ericsson has a proposal which could split the Open RAN alliance. We are starting to think they have a literal playbook with written instructions listing steps to take when confronted with open initiatives in the core domain.
A useful, readable primer about how network processors work. Good information, but we noticed they kept referring to products made by Lucent, and realized this piece is over 20 years ago. Holds up surprisingly well nonetheless.
Software and the Cloud
Earlier this month, Google sold off its well-regarded domain name service which sparked a surprising amount of vitriol online. One more product abandoned by Google. One person pointed out that this service was big enough that Google’s move risks calling into question its commitment to the rest of Google Cloud. We do not think this is in the cards, but if we were evaluating a multi-billion dollar cloud deal with Google we would absolutely factor their commitment into our decision-making.
Dropbox posted a fascinating blog post on how they select sites for their data centers. There is a lot of useful analysis in this, but worth pointing out that power is very much the limiting factor for data centers. Not just how much power the site will consume, but is there sufficient power in the region to support it. In many regions, the answer is no.
Contrary Takes
We have a few theses about the industry today. And while we are fairly certain they are right, we think it is worth considering alternative points of view when voiced well. Here are a few.
We talk a lot about the difficulties of porting software from x86 to Arm. This can be done simply with the click of a button in any modern compiler, but that only gets developers so far. There is a lot more work to be done to get the code optimized. This Twitter thread lists the low-level problems people face when porting code. It is fairly technical, but its obscurity is the point and shows why so many are perfectly happening foregoing the benefits of Arm.
A post on why one developer does not want to move to Arm CPUs, let alone RISC V. Summary - software matters and Arm and RISC V still have a long way to go.
Having non-x86 machines makes your life harder. Once upon a time we would have assumed this was an Intel-sponsored post, but we think this author is sincere, and while we do not entirely agree, he does raise some important points.
Other Interesting Items
A CPU video game that never quite made it out. If anyone wants to revive this, please sign us up for the Kickstarter.
How to steal art. File under back-up career plans.