Heating up a market for chips used in machine learning, Google LLC today cut prices for accessing the power of Nvidia Corp. graphics processing unit chips over its cloud service by up to 36 percent.

More specifically, Google said it will now charge 45 cents an hour to use Nvidia’s older K80 GPUs attached to a Google Computer Engine virtual machine, or computer emulated in software, and $1.46 an hour for the chip maker’s newer P100 devices, both in U.S. regions for now.

GPUs have become the chip of choice for deep learning, a method of helping a computer learn tasks from lots of data rather than being explicitly programmed, because they contain a large number of processing cores that can crunch data in parallel much faster than conventional processors. Deep learning has been responsible for recent breakthroughs in image and speech recognition, language translation and self-driving cars.

“Scientists, artists and engineers need access to massively parallel computational power,” Google Cloud Platform Product Manager Chris Kleban wrote in a blog post today. “Deep learning, physical simulation and molecular modeling can take hours instead of days on NVIDIA Tesla GPUs.”

Google is also lowering prices on so-called preemptible local solid-state drives that are attached to preemptible virtual machines, which can be interrupted and resumed later, by up to 40 percent, to 4.8 cents per gigabyte per month. Preemptible VMs, as Google described them in a blog post on another price-cutting move in August, greatly reduce compute costs on demanding tasks such as analyzing financial markets, processing data, rendering movies, analyzing genomic data, transcoding media and the like.

The timing of the price cuts is probably not random, since cloud leader Amazon Web Services Inc. is expected to announced a raft of new products, particular in artificial intelligence and machine learning, and price cuts next week at its annual AWS re:Invent conference in Las Vegas.

In late October, AWS launched a family of AI-optimized server “instances” based on Nvidia’s Tesla V100 chips, beating Google and Microsoft Corp.’s cloud unit to the market.

Google has trailed far behind AWS, Microsoft Azure, Alibaba Cloud and others because it got a late start in selling cloud computing services. But the company has vowed to turn Google Cloud Platform into a major player. Last week at the Structure cloud conference in South San Francisco, Urs Hölzle (pictured), senior vice president of Google’s technical infrastructure, said he still thinks Google cloud revenue could surpass the company’s massive advertising business, though not by the 2020 date he said a couple of years ago.

“We’re gaining market share actually fairly rapidly,” he said, though he conceded it’s still a relatively small buisness. Google has had some major customer wins this year, such as Salesforce.com Inc., Spotify, Snap Inc., but Hölzle said the big focus currently is large traditional enterprises.

Photo: Robert Hof