AI Firm Supported by Google and Nvidia Aims to Alleviate Chip Scarcity

Nvidia Corp. and a Google venture fund have participated in a seed funding round for a startup that aims to enhance the computing capabilities of specialized processors utilized in AI training. This investment has the potential to address a significant bottleneck in the rapidly growing field.

CentML, which builds software to help machine learning systems perform more efficiently, raised $27 million from investors including Google’s Gradient Ventures and Radical Ventures. Deloitte Ventures and Thomson Reuters Ventures also participated in the financing, the startup said in a statement.

The Toronto-headquartered startup aims to solve one of the biggest bottlenecks in AI development, the graphics processing units of Nvidia and its competitors, which process the massive amounts of data needed to train and operate AI systems. Supply may remain tight well into 2024, when prices will skyrocket, analysts predict.

Big-name backers are betting on young companies like CentML to find innovative ways around these constraints.

CentML was founded last year by Gennady Pekhimenko, a PhD in computer science from Carnegie Mellon University who is now an assistant professor in the Department of Computer Science at the University of Toronto. Pekhimenko and three others developed software that helped predict the time it would take to process tasks on different hardware. It monitors systems to pinpoint areas of underutilization—by analyzing costs, power consumption, and emissions—and then automatically allocates tasks to speed them up.

This, in turn, should help maximize the chip’s operating and shaving costs. The average utilization of GPUs in the market is around 30%, CentML said, citing research it conducted. Its technology can speed up systems “up to 8x, which will have a profound impact on our customers,” said Pekhimenko, also the startup’s CEO.

His startup now plans to open an office in Silicon Valley to attract talent. Pekhimenko’s goal is to double its workforce, now around 30, in the next 12 months.

“The size of AI models grew 10 times per year over the past decade, and the gap between computation and model size is widening,” he said in an interview. “There’s a desperation for computing, and chipmakers can’t deliver it fast enough.”