This article is more than 1 year old

Cerebras brings wafer-size AI chips to medical data analysis

CEO tells El Reg why biomedical firms dig big dies

AI chip startup Cerebras Systems has deployed one of its CS-2 systems at a well-funded startup that uses natural language processing to analyze massive amounts of biomedical data.

As announced on Monday, nference plans to use this CS-2 to train large transformer models that are designed to process information from piles of unstructured medical data to provide fresh insights to doctors and improve patient recovery and treatment. The CS-2 is powered by Cerebras' second-generation, Wafer-Scale Engine processor, so-called because the chip is wafer-size.

Cerebras said this deployment marks another significant customer win in the health care and life sciences space after installing similar systems at pharmaceutical giants GlaxoSmithKline and AstraZeneca as well as the US Department of Energy's Argonne National Laboratory for COVID-19-related research.

Andrew Feldman, CEO of Cerebras, told The Register this installation at Massachusetts-based nference is another testament to Cerebras' belief that its wafer-sized AI chips are better suited than traditional chips like Nvidia's GPUs for analyzing large amounts of data as fast as possible, which is increasingly important in areas like health care and the life sciences.

"All of this is profoundly computationally intensive. It is well suited for these new computational techniques and artificial intelligence. And these are exactly the techniques that we are hundreds of times faster [at] than [Nvidia]," he said.

In the case of nference, the Mayo Clinic-funded biomedical startup will use Cerebras' CS-2 system to train self-supervised learning models on large reams of unstructured medical data, which can include patient records, scientific papers, medical imagery, and genomic databases.

The unstructured nature of this data, which varies by file format, can be a massive headache for data scientists and machine learning researchers to process using traditional computing methods, according to Cerebras. The AI chip startup said this can even force researchers to resort to the boring and inefficient task of going through documents by hand.

"These are some of the most fertile research domains. They have vast amounts of data, and the human genome and other genomes are among them. They're unbelievably big," Feldman said.

What makes Cerebras's chips such a good fit for large data sets like this is their large size. While other semiconductor companies make several chips from dies cut from a single wafer, Silicon Valley's Cerebras makes one chip from an entire wafer. This allows Cerebras to pack a tremendous amount of processing cores — 850,000 in the case of its latest processor — on one chip that is made up of 2.6 trillion transistors.

Feldman said this makes Cerebras' Wafer-Scale Engine chip 56 times larger than the previous largest chip ever made, and it crucially allows large sets of data to remain on the chip during processing, greatly reducing the need for information to be repeatedly shuttled in and out, which takes time. He contrasts this super-die approach with Nvidia's architecture, which he claimed is slower because it relies on interconnects to ferry information between individual GPUs and CPUs when trying to perform the same level of processing.

"When we have to move it, we move it around the chip. And that's the fastest way you can move information on demand," he said. "So when you have to go off a chip to a Mellanox switch or to the CPU, which is what the GPUs need to do, you're somewhere between 1,000 and 10,000 times slower than if you can keep the information on a chip and move it on your piece of silicon."

What makes Cerebras' CS-2 system appealing to nference is the fact that it can accommodate longer sequence lines than traditional systems, according to Feldman.

"With the CS-2, they can train transformer models with longer sequence lines than they could before, enabling [them] to iterate more rapidly and build better and more insightful models," he said. ®

More about

TIP US OFF

Send us news


Other stories you might like