Graphcore IPUs ‘faster for AI-based drug discovery’ than GPUs
In the race to deliver the best machine learning accelerators, one of Nvidia’s main challengers has scored a victory in the biotech space, London-based LabGenius, which said intelligence processing units (IPUs) of Graphcore offers significantly faster performance for AI-based drug discovery than some unidentified mainstream GPUs.
Founded in 2012, LabGenius is a venture-backed company developing antibody treatments for cancer and inflammatory diseases by leveraging machine learning algorithms and lab automation to discover proteins that have the good qualities for treating medical conditions.
In a blog post to be published on Thursday, and seen by The register, Bristol, UK-based Graphcore, is set to reveal that LabGenius turned to its IPUs to train a BERT transformer model on a large dataset of existing proteins to predict masked amino acids. . This, in turn, we are told, has helped LabGenius identify important features of proteins that can help it develop new therapies.
We’ve reduced the turnaround time to around two weeks, so we can experiment much faster
Using Graphcore IPUs provided in cloud instances from Cirrascale Cloud Services, LabGenius said model training took about two weeks, down from the previous month with GPUs. As for GPUs, LabGenius didn’t say, but our best guess is that it was Nvidia, given the company’s dominance in AI, plus the fact that AMD is only start getting competitive in this area.
“Previously, we used GPUs and it took us about a month to have a working model of all the proteins out there. With Graphcore, we’ve reduced the turnaround time to about two weeks, so we can experiment much faster. , and we can see results faster,” said Dr. Katya Putintseva, Machine Learning Consultant at LabGenius.
LabGenius said getting a major performance boost like this is critical as it races to compete with other companies developing new treatments.
“Graphcore has changed what we’re able to do, accelerating the training time of our models from weeks to days. For our data scientists, this is truly transformative. They can move much more at the speed at which they think. For us, that’s incredibly valuable,” said Tom Ashworth, chief technology officer at LabGenius.
Nvidia watches Britain’s upstart Graphcore swing in the rearview mirror waving beastly second-generation AI chip hardware
Graphcore was founded in 2012 by semiconductor veterans Nigel Toon and Simon Knowles, and the chip designer believes its IPUs are better suited for AI workloads than GPUs because the architecture of the processors has been designed from the ground up with such applications in mind.
The chip designer said LabGenius plans to expand its use of standard PyTorch implementations of BERT provided by Graphcore, which require only small code changes. The biotech company is also looking to create new AI models using Graphcore systems, including graphical neural networks, where the chip designer said its IPUs have “an innate architectural advantage.”
Nvidia, Graphcore, and other AI chip designers, like Cerebras Systems, are all pursuing AI chip deals in the biotech space. For example, Cerebras announced last month that biomedical startup nference had adopted one of its CS-2 systems to analyze massive amounts of unstructured data. Nvidia, on the other hand, has developed entire software and hardware platforms for healthcare and life sciences companies. ®