Nvidia Groq 3 LPX claims 4x speed over Cerebras — but needs 64 chips to do it
nvidia's groq 3 lpx benchmark hit 3,400 tokens per second on gemma 4 31b, besting cerebras by 4x, but requires 64+ chips versus cerebras' one or two.
nvidia launched full production of the groq 3 lpx, an interactive inference accelerator built on the vera rubin platform to optimize token generation for ai agents. artificial analysis benchmarks show the system hitting 3,400 tokens per second on gemma 4 31b, but the figure requires splitting models across a rack of at least 64 lpux chips with 500mb memory each. the comparison to cerebras, which achieves 882 t/s with one or two chips, is contested due to this disparity in hardware scale and the exclusion of cerebras' newer cs-4 generation. despite the contested math, nebius plans to offer the chip via its token factory cloud service to address the latency needs of advanced agents.