Nvidia unveiled its Groq 3 NPU, an AI inference chip dubbed the Groq 3 LPU, at its GTC 2026 developer conference, SiliconANGLE reported on March 17 local time.
Unlike existing GPUs, the chip is optimized to run AI models. It is designed to complement GPUs by providing ultra-fast memory and low-latency performance for multi-agent systems.
The Groq 3 LPU runs in a dedicated server rack called the Groq 3 LPX, which is equipped with 256 units and supports 40PB/s of bandwidth, the company said. Nvidia plans to operate it with a rack called the Vera Rubin NVL72, which combines it with Rubin GPUs, to raise throughput per watt by 35 times and profitability by 10 times.
Nvidia also announced the Vera CPU rack, a Bluefield-4 STX storage rack and a Spectrum-6 SPX networking rack, in addition to the Groq 3 LPX and Vera Rubin NVL72.
Chi-gyu Hwang
delight@d-today.co.kr