Nvidia says Groq 3 LPX accelerators are in mass production and will reach market this year

Nvidia announced that Groq 3 LPX accelerators, acquired through a $20 billion deal in December, have entered mass production and are expected to reach the market before the end of this year. These chips are designed to accelerate inference tasks within AI infrastructure.

The accelerators will be integrated into Nvidia’s Vera Rubin platform and will be used for agent-based workloads. Nvidia also revealed that Nebius, a company founded by Arkady Volozh, has signed a contract to receive the Groq accelerators. Each server rack can house up to 256 LP30 accelerators, which are said to reduce latency in AI agent processing tasks.

Testing results shared by Artificial Analysis indicate that the Groq 3 LPX achieves a record throughput of 3,400 tokens per second with a context window of 100,000 tokens when running the open model Gemma 4 31B. Nvidia claims this delivers a fourfold improvement in system responsiveness compared to competing solutions, significantly cutting the time required for AI agents to complete tasks. The performance gains are attributed in part to the chip’s 500 MB of high-speed SRAM.

During the Hot Chips conference, Nvidia also disclosed a contract with SpaceX for supplying components of the Vera Rubin platform. SpaceX plans to use Vera’s central processors to coordinate operations between ground-based and space-based data centers. Samsung Electronics is manufacturing the Groq 3 LPX chips for Nvidia, as the original startup had partnered with the South Korean manufacturer, while Nvidia typically relies on TSMC for other products.

Nvidia’s founder, Jensen Huang, stated earlier this year that accelerators like Groq could capture up to a quarter of the data center market in the AI software development segment.

Reported by 3DNews.