GPU catalogue · NVIDIA
SRAM decode accelerator rack paired with Vera Rubin; in full production 2026-08.

Specifications
Good for: Very low-latency token generation.
| Machine | NVIDIA Groq 3 LPX |
|---|---|
| Memory | 500 MB SRAM per LPU; 128 GB SRAM + 12 TB DDR5 per 256-LPU rack |
| Architecture | Groq 3 LPU (LP30) |
| Status | Announced, expected 2026 |