Nvidia announces that Groq 3 LPX accelerator has entered mass production and will be available this year

December The $20 billion deal, which gives Nvidia access to the development work of startup Groq, has an unusual structure, but the company was the first to announce this week that three of Groq’s LPX accelerators are in production and will be on the market before the end of the year. These chips are specifically designed to accelerate inference in artificial intelligence infrastructure.

Image source: NVIDIA

Groq 3 LPX accelerators will be part of the Vera Rubin platform, and they will be used for agent tasks. The first Nvidia customer to sign a Groq accelerator contract is Arkady Volozh’s Nebius company, the founder of Yandex. Up to 256 LP30 accelerators can be combined in a server rack, which can reduce latency when processing AI agent task requests. Human analysis of the Groq 3 LPX test results in the open Gemma 4 31B model demonstrated shows that the accelerator delivers a record throughput of 3400 tokens per second with a context window size of 100,000 tokens. According to Nvidia, this improves system responsiveness by four times compared to competing solutions, significantly improving turnaround time for AI agents. This is largely facilitated by the presence of high-speed 500 MB SRAM inside the Groq 3 LPX die.

Meanwhile, at Hot Chips, Nvidia point outShe successfully signed a Vera Rubin platform component supply contract with SpaceX. Elon Musk’s company will use at least one Vera central processor to coordinate the work of data centers on the ground and in space. Nvidia’s Groq 3 LPX chips are manufactured by Samsung Electronics, as the original startup worked with the contract manufacturer. For the rest of Nvidia’s products, it’s more typical to use the services of Taiwan’s TSMC. Nvidia founder Jensen Huang said this spring that accelerators like Groq will capture a quarter of the data center market focused on using artificial intelligence for software development.