NVIDIA Unveils First Groq-Powered Chip LPX: Delivers Up to 35x Inference Efficiency Improvement per Megawatt with Vera Rubin Combo, and Showcases Next-Gen Kyber Prototype

According to 1M AI News's monitoring, Groq's 3 LPU (Language Processing Unit) is the first chip launched after NVIDIA's acquisition of AI inference chip startup Groq for around $20 billion in December last year, with shipments expected to start in the third quarter of this year. The Groq 3 LPX chassis can accommodate 256 LPUs, equipped with 128GB of on-chip SRAM and an extensible interconnect bandwidth of 640TB per second. The company claims that when LPX is deployed with Vera Rubin NVL72, the highest throughput per megawatt can be increased by up to 35 times, unlocking the revenue potential for trillion-parameter, million-token contextual inference scenarios. Huang Renxun described the two processors as "extremely different yet mutually unified: one pursuing high throughput, the other pursuing low latency," with LPX's on-chip memory significantly expanding the total memory capacity available to models. The LPX chassis is scheduled to be launched in the second half of this year along with the Vera Rubin platform.

During the conference, Huang Renxun also showcased the prototype of the next-generation chassis architecture, codenamed Kyber. Kyber will reconfigure the computation tray of 144 GPUs into a vertical arrangement to increase physical density, reduce latency, and will be integrated into Vera Rubin's successor platform, Vera Rubin Ultra, expected to be launched in 2027.