Finding what differentiates Cerebras' AI supercomputer

Finding what differentiates Cerebras’ AI supercomputer

With its newly released AI supercomputer, Cerebra Systems aims to give scientific research centers and other organizations a chance to experiment and see how the technology can apply to their goals and business model.

The AI ​​vendor — known for building computer systems for deep learning applications — on Nov. 14 introduced Andromeda, a 13.5 million core AI supercomputer. The supercomputer is built with a cluster of 16 Cerebras CS-2 systems and uses the vendor’s MemoryX and SwarmX systems.

MemoryX is a memory extension technology that Cerebras unveiled in 2021. It enables CS-2 to support models with up to 120 trillion parameters. SwarmX, also introduced in 2021, is a communication fabric that extends Cerebras Swarm on-chip fabric to off-chip fabric.

Andromeda does linear scaling on large language model workloads using data parallelism, distributing data in parallel computing environments.

Access for the public

As a startup founded in 2015, Cerebras’ differentiating factor is that its supercomputer is accessible to the public, said Karl Freund, founder and analyst at Cambrian AI Research.

“If you look at all the other competitors trying to muscle their way into this turf, they haven’t really set up something this large for public access,” Freund said.

If you look at all the other competitors trying to muscle their way into this turf, they haven’t really set up something this large for public access.

Karl FreundFounder and analyst, Cambrian AI Research

Cerebras competes with other AI hardware vendors, including much bigger Nvidia and fellow startup Graphcore.

And while organizations such as Argonne National Laboratory have bought architecture for supercomputers, they haven’t done so at scale, he said.

What Andrew FeldmanCEO of Cerebras, “has done here is realize that he needs to make the investment to create a supercomputer, give people access to it, let them find out what they can do with it that they cannot do with GPUs and let the chips fall ,” Freund continued.

Cost and partners

The biggest challenge for Cerebras’ AI supercomputer will be whether it is truly as easy to use as the vendor claims, Freund said.

Another challenge is cost.

“We know it’s very expensive because this is very exotic technology,” he added. “But you get a lot for your money.”

Early users of Andromeda include Argonne, which is using the technology to work on gene transformers, and Jasper, a vendor of AI marketing copywriting technology, which is using Andromeda to design its next set of AI large language models.

The benefit for a vendor like Jasper of using the AI ​​supercomputer is being able to experiment with large language models without spending the money needed to do it on their own, Freund said. “AI at this scale is still very much in phase of research,” he said.

Cerebras’ AI supercomputer also represents an emerging trend in the AI ​​hardware market of a rise in specialty accelerators, said Daniel Newman, analyst at Futurum Research.

“They’re purpose-built for the challenges like what Jasper AI is trying to accomplish,” Newman said. “Specializing accelerator chips is going to continue to be a critical component to the development of the next generation of AI.”

Leave a Comment

Your email address will not be published. Required fields are marked *