NVIDIA Groq 3 LPX
•256 x interconnected Groq 3 LPU accelerators (32 x 1U liquid-cooled trays, 8 LPUs per tray)
•AI inference compute: 315 PFLOPS (FP8) rack-level; 9.6 PFLOPS per LPU (FP8)
•Per LPU: 500 MB on-chip SRAM, 150 TB/s SRAM bandwidth, 2.5 TB/s scale-up bandwidth
•Total on-chip SRAM: 128 GB; Scale-up bandwidth: 640 TB/s
•Up to 12 TB DDR5 per rack
•Fully liquid-cooled, cableless design; built on NVIDIA MGX infrastructure
•Up to 35x higher inference throughput/MW vs HBM-based GPU inference