printed circuit board
getty
That's a really big chip.
A while back I wrote about Cerebras' “dinner plate” processor design. I think there are about 90,000 cores, and they are about 30 centimeters from corner to corner.
These huge pieces of hardware are just one aspect of a scalable design. Nvidia has Blackstone and its successor versions of GPUs. Some companies are looking at TPUs (Tensor Processing Units) and some are looking at the utility of ASICs.
But Cerebras has a niche in the industry.
On a broader level, what is happening has to do with the so-called “death of Moore's Law” and the inability of the community to reap predictable benefits from the doubling of transistors that has occurred over the past few decades. Now, this is a different game and parallelism is the name of the game. Innovators are changing the way hardware is manufactured. They are changing the way software is made. They are changing the way they deploy applications, and in the fast-paced days of next-generation foundational models, leaders are also thinking about reversing the massive sea change to the cloud to accelerate AI's “edge computing” achievements and encourage colocation to save resources. Oh yeah, the US is introducing nuclear power to its data centers.
The Cerebras dinner plate is, in a way, an omen.
What else is going on?
experts give their opinions
I participated in a panel discussion about this at Stanford University last year. This was interesting because one of the panelists, Michael James, who works at Cerebras as chief architect of Advanced Tech (he's also a co-founder), actually showed off one of the company's super chips.
“AI requires a lot of computing power,” he said in defense of Cerebras' strategy. And yes, he mentioned the death of Moore's Law.
“Many of us know this trend is coming to an end,” he said of the foresight of those close to the industry in advancing fundamental discussions about how to scale. “And I was like, oh, the field of computer science might be over. What are we going to do?”
But the apocalypse never happened, he pointed out.
“And really, our view is that[these trends]are just holding back innovation and creativity in this space. And as we're working on this kind of AI revolution and thinking about what intelligence is and what drives intelligence, if you asked people a while ago they'd say, 'These are formal rules of logic,'” James said. This is what it means to think and to be human. ”And if we look today, what we found is very different. We use these computers to solve optimization problems and find approximate solutions to unspecified equations, but that requires a lot more movement of information. ”
AI layers
Moderator Nina Gregory asked panelist Hilla Dangol about the six-layer approach applied to AI.
“We came across this layered approach where you introduce expected business outcomes as an output layer and service the business workflow,” Dangol explains, adding references to terms like sentiment analysis and fraud detection systems, as well as a term called “ML operations” before “LLM operations.”
“It's basically a data ML model, and it's mainly the output layer that I think made the six-layer framework so popular in my previous work,” he said. “When ChatGPT came out, we replaced it with ML operations, LLM operations. The structure hasn't changed except that we have one thing in common: the data.”
Computing for everyone
For Zyphra's Ilya Brown, democratizing technology is very important.
“At the heart of it, what we're thinking about is how do we make it accessible to everyone,” he told Gregory and the assembled audience. “So how do we make sure that everyone in the world has access to these advanced AI capabilities? A lot of what we have to think about is how do we think about cost reduction? How do we streamline some of these models? How do we make them more accessible on other types of chipsets and other types of hardware?”
The group then considered aspects of multimodal design that were of interest to the panelists, different types of challenges to progress, and other items. You can watch the entire conversation on YouTube.
Near the end, Gregory asked the panelists for their overall thoughts on where AI is headed.
“We're kind of ignoring computer science altogether,” James said, citing the fear of job losses and also noting the “huge challenges” facing humanity. “Just to get the information, you need some sort of log-time manipulation rather than a linear-time manipulation, and that requires a different silicon structure. The reason we have poor algorithms is because the memory is physically in a separate location from the processor.”
He also mentioned working with DARPA on next steps.
“There’s a world we’re building that aims to be a world that blends all of these things together,” he said.
In response to the same question, Brown said, “There are endless problems to solve.” “As we've seen here, we're thinking about everything from how we think about hardware and chipsets to how we think about different layers of the stack. I really encourage you to not only be passionate about your particular area, but also to show up regularly and see what other people are working on, because there's always something exciting.”
Dangor spoke about the value of power efficiency and energy-efficient systems and collaborative efforts among silicon designers, research scientists, and other stakeholders. He pointed out how the market will continue to demand the same results in the future. “It’s an added value for the consumer,” he said. “This is what all technology has given us so far.”
That's part of the context in which this hardware and software work is being done. As 2026 approaches, most of us could not have imagined where we are today even a year ago. It's dizzying, sometimes exhilarating, and sometimes frightening. Watch this space as we move into the new year.
