NVIDIA begins full production of Groq 3 LPX AI chip
What's the story
NVIDIA has announced that its Groq 3 LPX rack is now in full production. The move marks the commercialization of technology from the company's biggest acquisition ever. Back in December, NVIDIA had acquired assets from chip start-up Groq for $20 billion. The new racks will be deployed with Vera central processors and Rubin graphics processors at neocloud provider Nebius later this year.
Tech details
Each rack packs 256 Groq 3 chips
Groq's architecture features 500MB of on-chip SRAM, which helps reduce memory-related bottlenecks.
Samsung manufactures the Groq chips while Taiwan Semiconductor Manufacturing Company (TSMC) makes NVIDIA's GPUs.
Each LPX rack from NVIDIA packs 256 Groq 3 chips. According to a benchmark by Artificial Analysis, a single Groq 3 LPX rack can deliver up to 3,400 tokens per second.
Market competition
Low-latency inference
NVIDIA's move to manufacture and sell Groq chips highlights the rising demand for low-latency inference.
This is especially important in coding applications where users want quick responses without long delays.
The tech giant's strategy also puts it up against other companies targeting low-latency AI inference, such as AMD and OpenAI.
Sales forecast
Increase in Vera Rubin shipments
NVIDIA is also increasing shipments of its Vera Rubin systems, which went into production earlier this year.
At the launch event for Vera Rubin and Groq 3 LPX in March, NVIDIA CEO Jensen Huang had predicted $1 trillion in cumulative sales from Blackwell and Vera Rubin systems through 2027.