Column When I first saw an image of AI hardware startup Cerebras’ wafer-scale engine, my mind dismissed it. The company’s current product is about the size of an iPad and uses 2.6 trillion transistors contributing to more than 850,000 cores.
I felt it was impossible. Not a chip – not even a bunch of chiplets – but a monstrous silicon wafer.
Cerebras’ secret sauce allows it to treat that huge wafer as a single computing device, neatly connecting the bits that managed to make it through the manufacturing process, while the (no doubt numerous) bits that didn’t. are ignored.
That approach resembles the 17-year cicada strategy applied to computers: availability in such overwhelming numbers that it just doesn’t matter if thirty percent of the capacity disappears into the belly of every bird and lizard within 100 kilometers. Cerebras certainly says so – we’ll take the whole wafer, we’ll make as many processors as we can, and if a percentage of them don’t work, no problem.
At first it looked like Cerebras was just an outlier. But this year we learned that it was actually a precursor. When Apple introduced its Mac Studio — powered by its latest M1 Ultra silicon — we got a peek at what happens when a mainstream computer hardware manufacturer decides to go all-in: more than 100 billion transistors spread across a monstrous 1000mm.2 die. To cool down all that silicon, Apple had to strap a huge copper heat sink to the top of its chip. This monster breathes fire.
Oddly enough, the kind of monstrous computing power of M1 Ultra — and the new designs from Intel, AMD and Arm — hasn’t spawned a new class of applications that could take advantage of the hardware’s exponentially increased capabilities. We might have expected some next level artificial intelligence, augmented reality or some other game changer. The only thing Apple offered was the ability to edit 4K video more efficiently.
It feels like our ability to etch transistors onto a bit of silicon has outgrown our ability to imagine how to put them to work. The supercomputer of just 20 years ago has shrunk to a tidy desktop workstation, but where are the supercomputer-class applications?
In March, Nvidia announced its latest architecture – Hopper – with a preview of its merged ‘superchip’, featuring 144 Arm cores, integrated RAM and GPU. It basically squeezes a lot of data center functionality into a few huge chunks of silicon, all creatively bonded together to run efficiently. It’s another monstrous piece of hardware, but at least it offers a price-performance advantage for existing data centers.
On the desktop and in the data center, the pendulum is starting to swing back towards ‘bigger is better’. Moore’s law may still apply — until we run out of the periodic table — but my iPhone 13 is significantly bigger than my iPhone X, and my next MacBook Pro will be bigger and heavier than my 2015 model. Something has changed in computer science. While it makes sense that the data center should see such advancements, it remains unclear what that means for desktop and personal computing.
When I started my career in IT, mainframes had given way to minicomputers, and those minicomputers would soon be overrun by microcomputers. We have been living in the micro age ever since. Now we are witnessing the dawn of a new era: monster computers – devices where trillions of transistors and kilowatts of power fuse together.
But for what purpose? Raw capacity has never been the point of computing. These monsters need to be tamed, trained and put to work.