The arms race for accelerators that power machine learning tasks like facial recognition and voice processing is intensifying. Xilinx announced that it would buy Chinese startup DeePhi Technology to expand into software tools that run through the building blocks of machine learning more efficiently by mapping them onto the type of programmable chips Xilinx designs.
The deal dovetails with other machine learning efforts at the San Jose, California-based supplier of field-programmable gate arrays. Xilinx has partnered with Daimler to build driverless car systems based on its chips—more commonly known as FPGAs. Over the last four years, the company poured more than a billion dollars into the development of its new server chip architecture targeting big data and machine learning.
The deal could blow more wind into its sails. The company is locking down an engineering team with hard-to-find talent in both machine learning and FPGAs. Founded by researchers from Tsinghua University in 2016, DeePhi uses tools to boost power efficiency and lower the memory requirements of neural networks trained on reams of data—millions of miles of highway driving, thousands of photographs or hundreds of hours of human speech.
DeePhi maps algorithms into Xilinx FGPAs, giving them an edge over Nvidia’s graphics chips, which are the current gold standard for training neural networks but have less command of applying what they have learned—a process called inferencing. DeePhi can also reduce the number of computations required for inference, giving battery-powered surveillance cameras and drones the ability to run machine learning locally instead of within data centers.
“They also have an advanced understanding of high-level optimization of neural networks for size, performance and fit to FPGAs, especially using compression methods,” said Chris Rowen, former chief technology officer of Cadence’s IP business and founder of Cognite Ventures. “There are very few teams anywhere that understand model compression the way the DeePhi team does.”
The financial terms were not disclosed. But the deal underscores the mounting battle over chips that can power intensive machine learning tasks. Nvidia is reaping the richest rewards from the software industry’s shift to machine learning so far. But the company is spending billions to repulse threats from Google and other companies building custom accelerators to replace Nvidia’s chips in their data centers.
There are other challengers. Wave Computing and Graphcore are among the startups that have raised more than a hundred million dollars gambling on custom chips for neural-network training. Others focused on inference chips have also melted the hearts of venture capitalists long hardened against the semiconductor industry. Intel is spending lavishly to hold onto its data center dominance, which is under threat as Nvidia shifts the center of gravity there to accelerators.
Venture capital is also streaming into China’s semiconductor industry, which is focused on custom chips for facial recognition and autonomous driving. Many Chinese startups are targeting the country’s growing market for cars and surveillance cameras. Over the last two years, DeePhi raised more than $40 million in funding from investors including Samsung Venture Capital, MediaTek and Ant Financial.
Yet another investor was Xilinx, which has long depended on business from industrial, wireless, aerospace and other sectors where paying for new silicon every few years would be too costly. But under its new chief executive, Victor Peng, the company has doubled down on its growing business in data centers, where its chips are increasingly used as inference accelerators.
Despite lacking the mathematical might of graphics chips, the chips can perform simple, repetitive tasks very efficiently. But unlike graphics chips, FPGAs can be programmed in the field to conform machine learning algorithms, which are constantly changing. But that programming is still something of a black art—an issue that could hamstring FPGAs against rival inference technology, even GPUs.
“What will ultimately separate the winners is probably not their hardware, it is the ease of development and richness of available application experiences that run on that hardware,” Rowen said. “Xilinx has long had the kind of parallel computing fabric which is essential for good neural network implementation, but these tools have not been as mature as others in the industry.”
DeePhi gives Xilinx new tools to offer customers. The company has also crafted a custom architecture, called Aristotle, which runs convolutional neural networks mapped onto Xilinx’s chips. The architecture is currently used in products for surveillance cameras and servers but it can also be scaled down for smartphones. DeePhi’s other architecture, Descartes, runs recurrent and long short-term memory neural networks.
The deal could also improve Xilinx's chances against Intel’s Altera business, as more cloud computing companies like Amazon, Baidu and Alibaba install FPGAs in data centers rented out by businesses. Microsoft uses Intel chips as accelerators for its artificial intelligence services, though it plans to make them available to customers. Microsoft installs FPGAs inside every server added to its Azure cloud.
“Xilinx will continue to invest in DeePhi Tech to advance our shared goal of deploying accelerated machine learning applications in the cloud as well as at the edge,” said Salil Raje, vice president of the software and IP products group, in a statement. DeePhi, which has around 200 employees, will continue to be based in China.