Artificial intelligence chip startup Groq Inc. announced today that Nvidia will license its technology on a non-exclusive basis.
The deal will also allow the graphics card maker to hire several key Groq employees, including founding CEO Jonathan Ross and President Sonny Madeira. Alex Davis, CEO of Groq investor Disruptive Technology Advisers, told CNBC that the deal is worth $20 billion.
Through these deals, companies acquire the talent and technology of startups without making a direct purchase, a practice known as a reverse takeover. They avoid the antitrust scrutiny common in acquisitions. Over the past three years, tech giants such as Microsoft and Meta Platforms have signed several such agreements to advance their AI roadmaps.
Nvidia will reportedly pay $20 billion to license Groq's intellectual property, a premium of $13.1 billion over the company's September valuation. The two companies stated that the license covers Groq's "inference technology," presumably referring to its flagship LPU inference chip.
The company claims that the LPU can run inference workloads with ten times less power consumption than a graphics card. According to Groq, the chip's efficiency stems from its deterministic design. This means it can control computation timing with high precision. In contrast, standard non-deterministic chips often suffer from unexpected processing delays, leading to disordered computation completion.
The LPU also optimizes AI performance in other ways. It contains hundreds of megabytes of on-chip SRAM, the fastest type of memory on the market. This memory pool outperforms the HBM memory used by graphics cards and consumes less power.
Groq connects LPU-equipped servers into inference clusters through its internally developed interconnect system, RealScale. According to the company, this technology solves a problem known as crystal-based drift, which makes coordinating AI servers difficult.
Processors use clocks to control the frequency at which circuits perform computations. Typically, clocks are implemented using a tiny quartz crystal. Crystal-based drift is a phenomenon that causes clocks to unexpectedly slow down, leading to inefficiencies in the AI ??inference process. Groq states that RealScale can automatically adjust the processor clock to mitigate this problem.
According to reports, Nvidia CEO Jensen Huang informed employees in an internal memo, "We plan to integrate Groq's low-latency processors into Nvidia's AI Factory architecture, extending the platform to a wider range of AI inference and real-time workloads."
Following the transaction, Groq will continue to operate as an independent company. Chief Financial Officer Simon Edwards will succeed Jonathan Ross as CEO.
As of July, Groq reportedly projected year-end revenue of $500 million. The company sells access to its chips through a cloud platform called GroqCloud. The platform also offers open-source AI model libraries and tools, such as a search engine, which can be used to process cues.
Tags：