Nvidia says its Groq 3 LPX is four times faster than Cerebras, but the math is more complicated
decoder ·
Nvidia has begun full production of its Groq 3 LPX inference accelerator, designed for ultrafast token generation in agentic AI systems. An independent benchmark recorded 3,400 tokens per second on the Gemma 4 31B model, which Nvidia claims is four times faster than Cerebras's 882 tokens per second. However, experts argue the comparison is stacked in Nvidia's favor because the LPX system required at least 64 accelerators, while Cerebras used only one or two, and the newest Cerebras CS-4 generation was not included. The Groq architecture relies on SRAM-heavy dataflow chips with only 500 MB of memory per LPU, far less than a Rubin GPU's 288 GB, forcing models to be split across racks of up to 256 LPUs over Ethernet. This setup works well for dense models like Gemma 4 31B that fit in one rack, but larger mixture-of-experts models like DeepSeek V3 would need over 1,300 accelerators. Nvidia acquired the Groq license for about $20 billion, and Nebius plans to be the first cloud provider to offer the chip through its Token Factory.
decoder