Nvidia Acquired AI Startup That Shrinks Machine-Learning Models
THE TAKEAWAY
• OmniML helps complex models run on devices
• The deal could help Nvidia improve its chip for autonomous vehicles and robots
• Nvidia faces competition from smaller chipmaker in edge computing
Nvidia in February quietly acquired OmniML, a two-year-old artificial intelligence startup whose software helped shrink machine-learning models so they could run on devices rather than in the cloud, according to a spokesperson and LinkedIn profiles of former OmniML employees who now work at Nvidia.
The acquisition could be a sign that the chipmaker, whose data-center server chips have fueled a recent AI boom and enabled chatbots including ChatGPT, wants to improve its separate AI chips for cars, industrial robots and drones. The startup’s engineers also could potentially aid an effort to shrink the AI software that powers chatbots so it can run on devices rather than in data centers.
Terms of the deal weren’t disclosed. Since the acquisition, Nvidia has become an increasingly active investor in hot AI startups that buy its chips. In recent weeks, for instance, it bought equity stakes in chatbot developer Inflection AI and video-editing software firm RunwayML. (See our Generative AI Database.)
Nvidia designs most of the world’s graphics processing units, which have been particularly useful for developing AI software such as large-language models, the technology behind ChatGPT.
Now developers of LLMs, including ChatGPT creator OpenAI, are increasingly working on ways to make the software smaller so it can run on industrial or personal devices such as smartphones or laptops. Running such software locally on a device is typically faster and less expensive than running it in the cloud.
It isn’t clear whether the OmniML technology or personnel would aid in the effort to shrink LLMs. But the LinkedIn page of one of the startup’s founding engineers says he now works at Nvidia on “model and pipeline optimization for large generative models.”
The startup previously said its software could compress the size of machine-learning models so they could power AI on devices, but the examples it cited primarily involved computer vision—helping a smart camera or autonomous vehicle identify objects around it, for example.
The OmniML technology would fit in with Nvidia’s existing business of selling chips that process data on vehicles and industrial devices so they can handle such tasks. Nvidia does not break out financials around that business, known as edge computing.
Nvidia has cornered the market on GPUs for data center servers, but it faces some competition in edge computing chips.
Startup SiMa.ai, for instance, is working on AI chips for devices ranging from robots to cars to cameras. The startup has raised $200 million in venture capital and recently beat Nvidia in a closely watched test, called MLPerf, of chip power and performance.
Nvidia declined to comment on when the OmniML purchase closed and how much it paid to buy the startup. OmniML said it raised $10 million last year in a seed round led by GGV Capital that also included Qualcomm Ventures and Foothill Ventures. In January, OmniML announced a strategic partnership with Intel, an Nvidia rival, to bring its software to Intel hardware for some customers.
That partnership may have ended following the Nvidia purchase.
Intel did not immediately respond to a request for comment.
OmniML was founded by Song Han, an electrical engineering and computer science professor at the Massachusetts Institute of Technology; Di Wu, a former software engineer at Meta Platforms; and Huizi Mao, who co-invented “deep compression” technology, which came out of Stanford University and is at the heart of the startup’s software.
Wu and Mao, who did not immediately respond to a request for comment, now work at Nvidia, according to their LinkedIn profiles.
“AI is so big today that edge devices aren’t equipped to handle its computational power,” Wu said in a March 2022 press release.
Nvidia plans to use OmniML’s technology to help Nvidia’s customers develop AI models faster, including by improving accuracy and reducing latency for complex machine-learning models, according to the spokesperson.
The acquisition happened three months before Nvidia stock jumped 25% following its announcement of a blowout sales forecast for the current fiscal quarter, which ends in July, owing to exploding interest in LLMs.
Nvidia has a market capitalization of $1 trillion.