The Groq 3 LPU chip (~500 MB on-chip SRAM, ~150 TB/s bandwidth), deployed as the LPX rack of 256 LPUs (~128 GB aggregate SRAM), takes the decode/FFN slot in Vera Rubin serving (no HBM); Rubin GPUs keep prefill and attention. NVIDIA announced Groq 3 LPX in full production on 2026-08-24, with Nebius the first AI cloud to adopt it. NVIDIA holds a non-exclusive licence to Groq's inference technology (December 2025, about $20B reported). Rubin CPX, announced 2025-09-09, was pulled at GTC 2026 to prioritize LPU decode.
LPX in full production, 2026-08-24observed
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | Performance figures for LPX are NVIDIA's (vendor claims). Groq remains an independent company; its own cloud business is separate from NVIDIA's licensed LPX product. The LPU runs on a compiler-scheduled backend that NVIDIA coordinates with GPUs through Dynamo. |
| Caveat | At GTC 2026 NVIDIA's Ian Buck said CPX was pulled to prioritize LPU decode and could be reconsidered for a later generation. As of 2026-10-16 NVIDIA had announced neither a permanent cancellation nor a revised delivery date. On 2026-08-31 analyst Ming-Chi Kuo reported, and TechNews relayed on 2026-09-02, that NVIDIA had brought CPX back as a 168 GB HBM4 part in a separate rack, with production in Q1 2027. NVIDIA has not confirmed the report. |
| As of | 2026-10 |
| Source | NVIDIA Technical Blog, Inside NVIDIA Groq 3 LPX (2026-03-16); NVIDIA press release (2026-08-24); Tom's Hardware transcript of NVIDIA press Q&A (2026-03-23) · LPX chip and rack specifications; full-production announcement 2026-08-24 |
| Review | checking…review by 2027-01-31 · standard cadence |
| Recorded changes | last 2026-10-15 · 4 revisions tracked |
| Claim id | groq-3-lpx-prefill-sram-lpu-fills-the |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.