D2D: Accelerating Towards Chaos
A deep look into Intel's woes, an update on RISC V and SiFive and grasping at the AI accelerator landscape
In which we provide a series on Intel’s current difficulties and make a few suggestions on the path forward. We also examine the growing competition among AI accelerators, with multiple start-ups chasing a limited market and then explore the reasons that got us here and make a few suggestions on the path forward.
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 reported some pretty ugly results. This has created a significant credibility gap with the Street. At this stage, we think the best thing the company should do is arrange for a syndicate of strategic investors - aka Intel Foundry customers like Qualcomm and Broadcom - to buy a minority stake in the company. There is precedent for this, ASML did something similar 10+ years ago to fund development costs for EUV. This would mark a bottom for the stock and send an important signal.
For the first time, we have investors asking us if Intel can survive. By the numbers it looks challenging. No company that has fallen off the Moore’s Law curve has ever managed to climb back on the train. Intel needs to iron out many problems to get there. Chief among our concerns is the way in which the company seems to have been surprised by factors that were already fairly well known. For instance, they should have ended the dividend two years ago. By our estimate, the company now faces six to eight quarters in the wilderness - during which product results will likely continue to disappoint while the company is racing to get manufacturing process into volume production. Now is the time for the company to align on a single message stream with a complete overhaul of its next two years of public appearances.
It has been a while since we have written about RISC V. So the news that SiFive has launched a new product targeting datacenter workloads gave us a chance to look at what has been happening in the RISC V ecosystem. The community is progressing, but has entered a ‘boring’ phase - important but full of unglamorous working optimizing software to run on RISC V chips.
Lately there have been a number of prominent fund raises from start-ups developing AI accelerator chips. As much as we are excited to see US venture funding for semis, we think this segment is going to prove very difficult. Put simply, there are two many companies chasing too small a market - with Nvidia at the top, AMD and Intel fighting for a toehold, and all the major customers designing their own chips. This does not leave much for anyone else. The obvious question is how did we get here? We think the problem rests with two decades of US venture under-investment in semis. The big VC funds have lost the institutional knowledge of the sector and have fallen back on pattern matching resulting in too many very similar companies. There are much more promising semis investment areas out there, but it will probably take a new generation of venture funds to unlock that value.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
Intel’s Immiseration. A good run down on the problem and full credit for the title.
Lost in the barrage of bad news, Intel reached an important milestone - version 1.0 of their PDK. Their new 18A manufacturing process is now ready for customers.
Google published a retrospective on 10 years of TPUs. Nothing too surprising here, but a good reminder that Google is good at this and its hyperscale competitors are years behind.
Another RISC V data center processor supplier has decloaked. Interesting timing, right on top of SiFive’s announcement. Aleena has over 100 employees, which is a big team.
Meta posted a paper on its Llama 3 model, and buried deep inside that are reports of just how hard it has been maintaining a large H100 cluster. The industry is going through growing pains with this new form of compute. There is a surprising amount of friction in deploying GPUs, even highly seasoned teams (and Meta is among the best) are encountering considerable difficulties here.
Three people (across four companies) reportedly control a half million Nvidia GPUs. There are many ways to read this, our take is that this is not natural and probably spells trouble for the inventory cycle somewhere down the road.
Ampere launched a 512 core CPU. At some point the distinction between a CPU and GPU is just going to blur into insignificance. Ampere has been noticeably quiet lately, they face a lot of market challenges (see above), so we are glad to see they are still pushing the product envelope.
Networking and Wireless
Verizon appears to be preparing to sell off more of its cell towers. This has always struck as a triumph of financial engineering over sound strategy. The big US carriers get rewarded for exiting these asset-intensive holdings, but somehow the new owners attract very healthy valuations, which seems to imply there is a lot of value lost to the operators. To say nothing of how losing control of the towers affects future deployment plans.
Remember when Cisco was the dominant player in networking, a high-flying growth company? Probably not, those days were long ago. The company just layed off another 4,000 people as it pivots to focus on AI and Security. Our sense is that the market has just moved away from them, and now the company is transforming into a legacy licensor of long-lived infrastructure software like Oracle. This is not a ‘bad’ thing, just a natural evolution. Campus switches, once Cisco’s core product, are now a steady-state low growing category. The market moves on and new models emerge.
Software and the Cloud
One of the areas where advocates have promised big advances from AI is in the field of drug discovery and biological chemistry. This piece argues, we think correctly, that the gains from AI have already been achieved here - beyond this we start to run out of data. We think there is a pattern emerging here - AI has so far helped to advance things that we were already doing (AI as a feature). Beyond this we are going to need a combination of more data (as argued in that piece in conjunction with new lab work), using AI to generate that data (which probably only works in a few fields), and some unforeseen breakthrough.
Nathan Egge from Google provides a good overview of the status of RISC V software development.
Developing mobile apps has become more complicated. The development process is now burdened by an increasingly complex software ecosystem and more critically new requirements from the platform owners Apple and Google.
Diversions
How to design biological circuits. We are not sure of the utility of these, but it is fun to think about.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.