Nvidia Launches First Groq LPX Chip: Up to 35x Improvement in Inference Efficiency Per Megawatt When Combined with Vera Rubin, and Demonstrates Next-Generation Kyber Prototype

Gate News: On March 17, NVIDIA launched the Groq 3 LPU (Language Processing Unit) chip. This is the first chip released after NVIDIA acquired AI inference chip startup Groq for approximately $20 billion in December last year. Shipment is expected to begin in the third quarter of this year. The Groq 3 LPX rack can hold 256 LPUs, equipped with 128GB of on-chip SRAM and an interconnect bandwidth of 640TB per second. Officially, when deployed with Vera Rubin NVL72, LPX can increase inference throughput per megawatt by up to 35 times, unlocking revenue potential for trillion-parameter and multi-million-token context inference scenarios. Jensen Huang described the two processors as “extremely different yet unified: one pursuing high throughput, the other low latency,” with LPX’s on-chip memory significantly expanding the total available model memory capacity. The LPX rack is planned to be launched later this year alongside the Vera Rubin platform. Additionally, Huang also showcased a prototype of the next-generation rack architecture codenamed Kyber. Kyber rearranges 144 GPU compute trays vertically to improve physical density and reduce latency, and will be integrated into the successor platform Vera Rubin Ultra, expected to be released in 2027.