Source item

Nvidia’s Groq 3 LPX-based server racks have entered mass production at Foxconn, the primary supplier of the racks, media report, adding Nvidia announced the Groq 3 LPU chip in March. Samsung Foundry is producing the 4nm chips, which feature 500MB of on-chip SRAM per chip, better for inference (decode) and sidestepping the memory chip bottleneck. The Groq 3 LPX rack combines 256 LPUs for 128GB of total on-chip SRAM. 1/2 $NVDA $SSNLF #Foxconn #Samsung https://t.co/jCByuvICBa 2/2 Nvidia Groq 3 LPX servers require a different build due to rapid token generation and high frequencies, leading to some high-end material upgrades, media report: -High-end CCL (Copper Clad Laminate) -HVLP4 low-roughness copper foil (High-Very-Low-Profile-4) -M9 CCL paired with Q-Glass (quartz fabric) -M8 CCL paired with Low Dk2 glass cloth, and M8+ with Low Dk3 -Q Glass from Hubei Feilihua Quartz and Glotech (5475) -Low Dk3 glass cloth from Nittobo (3110) and Fulltech (1815) -HVLP4 Copper Foil from Mitsui Mining (5706) and Co-Tech (8358) -M9 CCL from Elite Material Co. (EMC, 2383) $NVDA https://t.co/5ozRghvnKt