Qualcomm has had much experience designing small form factor low-powered chips for handheld devices. Now it has banked on that expertise to deliver a compact but fast AI-accelerated low-power chip destined for use in huge data centers.

Mike Dano writes about Qualcomm’s new fast chips in this article from LightReading:

“The market for neural net inference acceleration is set to take off as machine learning features are incorporated into more cloud applications,” Kevin Krewell, principal analyst at Tirias Research, told Light Reading in an email. “To fit into more easily into existing rack servers [in data centers], these new inference cards need to be low power and compact in size. Qualcomm is using its experience in low-power mobile designs to build a very power efficient accelerator, with a very competitive software stack.”

Qualcomm, of course, built its business developing the technologies and silicon for many of today’s smartphones. Such devices require tiny chips that don’t use much power — attributes increasingly prized by data center operators.

Further, for more than a decade Qualcomm has been working to insert AI capabilities into its chips for phones, and now it is working to leverage those efforts with a new AI acceleration chip for the data center market. As Tirias’ Krewell pointed out, Qualcomm’s new Cloud AI 100 chip will be competing against the likes of Nvidia’s Tesla T4 PCIe cards and Intel’s Nervana chips, as well as products from startups like Habana Labs.

Qualcomm’s new data center chipset is part of a wider push by the company to break into new areas for growth — an effort that’s taken on new importance following Qualcomm’s failure to consummate its proposed merger with NXP. Qualcomm had been betting on that transaction to expand its addressable market into the broader automobile industry, and now as a result Qualcomm is working to stamp out new business beyond the cooling mobile phone sector with organic efforts like the Cloud AI 100.

As for the Cloud AI 100 specifically, Qualcomm is boasting that the product represents a 10x performance over the combinations of CPUs, GPUs and/or FPGAs that are used in today’s data centers. The company added that the market for such chips could grow to a $17 billion opportunity by 2025. The company said the chip uses the 7nm process and can support AI systems including PyTorch, Glow, TensorFlow, Keras and ONNX.