Cerebras Systems unveils a 1,200 trillion processor of slice scale transistors for AI

This site may generate affiliate commissions from the links on this page. Terms of use.

Credit: Getty Images

The number of modern processor transistors is huge – AMD announced earlier this month that a full implementation of its 7 nm Epyc "Rome" processor weighed 32 billion transistors. For that, Cerebras Technology says, "Hold me right." The AI-based company has designed what it calls a Wafer-scale engine. The WSE is a square about eight inches out of nine and contains about 1.2 trillion transistors.

I'm really surprised to see a company market a product the size of a wafer as quickly. The idea of ​​scaling scale treatment has recently attracted attention as a potential solution to performance scaling issues. In the study we talked about earlier this year, the researchers evaluated the idea of ​​building a huge GPU on most or all of a 100 mm wafer. They found that the technique could produce viable and high performance processors and that it could also efficiently adapt to larger nodes. The Cerebras WSE really qualifies as Lorge wide – its total area is much larger than the hypothetical designs we envisioned earlier this year. This is not a 300 mm wafer of normal size, but a surface greater than that of a 200 mm.

The biggest GPU, just for comparison, measure 815 square millimeters and 21.1B packs transistors. So the Cerebras WSE is just a bit bigger, as these things are going. Some companies send pictures of their chips next to a very small common object, like a quarter. Cerebras sent a photo of their dice next to a keyboard.

As you can see, it compares quite well.

Cerebras WSE contains 400,000 sparse linear algebraic cores, 18 GB of total on-die memory, a memory capacity of 9 Pbps on the chip and a distinct bandwidth of up to 100 Pbps. The entire chip is based on TSMC's 16-nm FinFET process. Since the chip is built from (most) of a single wafer, the company has implemented routing methods around faulty cores and can keep its bays connected even though it has bad cores in it. part of the plate. The company claims to have redundant cores implemented on chip, although it has not discussed details yet. Details on the design are presented at Hot Chips this week.

The WSE – "CPU" just does not seem to be enough – is cooled with a massive hot plate resting above the silicon, with vertically mounted water pipes used for direct cooling. Cerebras has designed its own software, because there is no traditional housing large enough. PCWorld describes it as "combining a printed circuit board, the wafer, a custom connector connecting the two and the cold plate". The details of the chip, such as its raw performance and its energy consumption, are not yet available.

A fully functional scale – scale processor, marketed on a scale, would be an interesting demonstration of the relevance of this technological approach to the wider market. Although we will never see consumer components sold in this way, we are interested in the use of platelet-level processing to improve performance and energy consumption in various markets. If consumers continue to move workloads to the cloud, especially high-performance workloads such as gaming, it's not foolish to think that GPU manufacturers will take advantage of this idea and create sets of components that no one could afford to feed into the cloud. gaming systems in the future.

Now read: