{"id":109305,"date":"2026-04-19T01:36:23","date_gmt":"2026-04-19T04:36:23","guid":{"rendered":"https:\/\/mastertrend.info\/?p=109305"},"modified":"2026-04-23T16:20:19","modified_gmt":"2026-04-23T19:20:19","slug":"groq-3-lpu","status":"publish","type":"post","link":"https:\/\/mastertrend.info\/en\/groq-3-lpu\/","title":{"rendered":"Groq 3 LPU and Nvidia's new inference strategy"},"content":{"rendered":"

Groq 3 LPU and the strategic shift at Rubin<\/h2>\n

The unveiling of the Groq 3 at GTC 2026 is more than just a technical launch: it marks a strategic shift in how Nvidia structures its inference platform. More than just a new chip, it redefines Rubin's internal hierarchy and anticipates a distinct phase in the competition for specialized silicon.<\/p>\n

At GTC 2026, held in San Jose, Nvidia unveiled the Groq 3 inference accelerator: the first chip to emerge from its $20 billion licensing and talent agreement signed on December 24, 2025. It is an LPU (language processing unit) based on SRAM that Nvidia<\/a> It's integrated into the Vera Rubin platform as a dedicated coprocessor for the decoding phase. The manufacturer announced an expected shipment date of the third quarter of 2026; production will be handled by Samsung on a 4nm node. It's also Nvidia's first rack-scale product designed around non-GPU silicon, and its arrival has prompted a reordering of its own components in the roadmap.<\/p>\n

The heart of the Groq 3 LPX is the LP30 chip: 512 MB of SRAM per die and 150 TB\/s of memory bandwidth per chip. To put this in perspective, a Rubin GPU with 288 GB of HBM4 offers around 22 TB\/s; the order-of-magnitude difference is not a nuance but an architectural choice. A full LPX rack houses 256 LPUs, totaling 128 GB of SRAM and 40 PB\/s of aggregate bandwidth. Nvidia claims that, combined with a Rubin NVL72, an LPX rack delivers up to 35 times the performance per megawatt compared to an NVL72 alone in trillion-parameter models, with an operating cost target of $45 per million tokens.<\/p>\n

Groq 3 and the function in Rubin<\/h2>\n
<\/picture> \"Rubin<\/picture>
Nvidia outlined its seven-chip Rubin SuperPOD strategy at GTC 2026. <\/span>(Image credit: Nvidia)<\/span><\/figcaption><\/figure>\n

In the planned operation, Rubin GPUs handle the prefill phase\u2014processing long contexts and high-density calculations\u2014while Groq LPUs manage decoding and token generation with reduced latency. Dynamo orchestrates this heterogeneous distribution, assigning tasks based on batch size and parallelism to balance performance and energy cost.<\/p>\n