D2D: Intel's New Reality
Ten things we hate about Nvidia and Qualcomm's growing developer outreach
Highlights from out blog including a look at Intel Foundry’s numbers, the present and future of Intel, Nvidia’s potential weaknesses and the case for Deep Tech Venture funding with the rise of AI because all that AI has to run on someone’s hardware.
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 released the financials for its foundry Intel Foundry unit (IF). We made a stab at these a few months back and we were in the ball park, but the Street did not like the numbers. The timeline for IF reaching steady state margins is 2030, again something we have been saying for a while, but beyond the average tenure of current investors, It is time to start thinking about Intel for what it is, not for what it once. They are never going to be the 90% share, king of the data center again, the world has just changed too much. As the Boss says, No one cares about your glory days. Instead, they are should be appreciated for being one of the leading players, with immense technical talent and channel relationships. That is not a terrible position.
Qualcomm is re-learning how to evangelize their products to developers. Spurred by the imminent launch of Snapdragon-based PCs, as well as the emerging opportunity to drive AI at the edge, they have made some notable strides in building up support for third party developers to access Qualcomm chips and software libraries. We still have many questions about the company’s long term position, but they can still execute well when they put their minds to it, and those efforts are starting to look fairly solid on this front. Now they just need to bring back the Uplinq conference.
Nvidia has all the momentum right now, so we set out to look for the ways in which it could be vulnerable. They have a clear lead in hardware and a massive lead in software, but there is still room for competitors to at least build toeholds. Nvidia also risks becoming a victim to its own success as its growing power pushes customers to alternatives
We remain big believers in the opportunity for Deep Tech for venture investors. The rise of AI reinforces that thesis, as it opens doors to new models of compute and new business models for hardware companies.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
Doug O’Laughlin has a smart take on how to think about Nvidia’s approach to systems architecture. He holds that the company is finding ways to advance compute density in the wake of Moore’s Law slowing, by focusing on the architecture of entire data centers. Not just a chip, not a just a server or a rack, but the whole data center.
Austin Lyons of ChipStrat on the depth of Nvidia’s software moat.
The CEO of Groq gave an interview where he claimed the company no longer sells hardware, focusing solely on their Inference as a Service offering. Their business model attracts a fair amount of criticism, but we think it is a sensible move, if their hardware is as good as their benchmarks say.
AMD open sourced its ROCm software. Open source is probably the only way to tackle Nvidia’s Cuda lock. Of course there is open source and “open source”.
TinyBox has a self-proclaimed mission to bring AI compute to average users. Over the last few weeks they stopped selling the AMD version of their hardware because of a lack of firmware support for their software. They eventually reversed course, but this is a highly visible example of AMD’s problems climbing over Nvidia’s CUDA moat.
Everyone is now convinced that we are on the cusp of massive advances in automation of industrial systems. This author details why this has not happened in mining already. His take is that industrial systems are optimized for humans, and adding a little bit of automation can actually make things worse. To really automate these systems they will need to be entirely rebuilt for robots.
Networking and Wireless
A few years ago, Private Cellular Networks were the hot topic in telecom. Fast forward to today, and the enthusiasm for the technology has waned, at least among market forecasters and their telco customers. It turns out these are useful, but not in the ways the marketing departments originally thought.
A catalog of Android malware. The sophistication of attacks against iOS is definitely growing, but Android remains the target of choice for much of the malware universe.
The network industry is messy right now. Like everywhere else in enterprise electronics, the weight of the hyperscalers has warped the market, router sales have flat-lined as data center switches adopt many routing functions, at least the ones the hyperscalers need. A big part of this is the current urgency of using Infiniband (from Nvidia) for its low latency properties in AI workloads. The networking majors, like Arista, are promoting their version of low-latency Ethernet, UltraEthernet, but that is not quite ready, and Nvidia probably outsells Cisco and Arista in the data center right now.
Software and the Cloud
AI consciousness is inevitable, say these Carnegie Mellon researchers. Yann LeCunn, head of Meta’s AI team, and a heavyweight in the field says “LLMs will never reach human level intelligence.” So that’s totally clear, glad we have resolved that topic.
Apple is rumored to be unveiling on-device generative AI that “outperforms GPT-4”. We do not usually spend much time on Apple rumors, but we think this highlights the way in which ‘old’ models like GPT-4 are getting rapidly deflated to run on mobile devices and the edge.
Expect to hear a lot more about Retrieval Augmented Generation (RAG) as an alternative to large language models (LLMs). The two are not actually in conflict, and will likely ended up being used in conjunction with each other. Very few of us will really need to understand what RAGs do, but they perform perfectly well on CPUs, so certain vendors are going to promote their use. Chip vendors not named Nvidia are going to heavily promote this line of reasoning.
3Blue 1Brown provides a good explanation of how transformers work - deep enough to convey the nuances, high enough that anyone can understand.
We have linked to a few stories about VMWare customer complaints in the aftermath of the Broadcom acquisition. Broadcom has responded, with one executive saying the new pricing scheme will actually save some customers money. Knowing Broadcom’s history, we think VMWare will end up more focused on a smaller number of large customers, and not worry too much about losing the long tail.
Diversions
Englishman runs 10,000 miles, the length of Africa.It was fun to watch his progression on TikTok.
Mexican academics discovered a new trove of very early documents dating back to the late days of the Aztecs and early days of the Spanish conquest. These accounts, often firsthand, can greatly expand our understanding of pre-Columbian Mexico. Not as ancient as the Herculaneum scrolls, but equal potential to shed new light. The Fifth Sun, which narrates much of that period via previously discovered codices is one of our favorite history books.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.