NVIDIA unveils Vera Rubin architecture and Groq 3 LPX at Hot Chips 2026
NVIDIA is set to unveil its next-generation AI infrastructure at Hot Chips 2026, centering on the Vera Rubin platform and a new partnership with Groq. The upcoming system moves beyond isolated chips to present a cohesive hardware ecosystem designed for massive-scale AI factories.
The core of this architecture remains the Vera Rubin NVL72, which will handle large-scale context processing, training, and inference. To address latency-critical tasks like token generation and decoding phases, NVIDIA is integrating Groq 3 LPX racks. Each system can house up to 256 LP30 accelerators coupled via direct chip-to-chip connections, delivering a reported output of 3,400 tokens per second for the Gemma 4 31B model.
Network scalability is being addressed through Spectrum-X Multiplane, which allows scaling up to 512,000 GPUs without adding a third switching layer. This design maintains roughly 90 percent of total bandwidth even if one plane fails, with hardware-based recovery reportedly eleven times faster than software balancing. The switches utilize a 102.4 Tbit/s Spectrum-6 ASIC.
Supporting these components are BlueField-4 for network and storage management via DOCA, and NVLink Fusion, which combines sixth-generation NVLink with C2C technology to connect custom XPUs and CPUs. NVIDIA also highlights the ConnectX-9 SuperNIC, supporting up to 1,600 Gbit/s per GPU when paired with Spectrum-X switches.
Гика
