Tenstorrent Previews Large Compute Cluster, Generates Video Faster Than Real Time

Tenstorrent has unveiled its next-generation cluster-scale computing systems, showcasing a video generation demo ahead of its full launch. The system, powered by an optimized version of the Wan2.2-14B... Read original →

Tenstorrent’s BlackHole-powered cluster achieving sub-real-time video generation triggers a cascade: EDA vendors like Siemens and foundries like GUC must accelerate RISC-V heterogenous support and high-bandwidth die-to-die interfaces. By avoiding proprietary interconnects, it sidesteps NVIDIA’s ecosystem lock-in but invites EU AI Act scrutiny on generative model transparency, potentially inflating software compliance costs. While NVIDIA doubles down on fixed-function Transformer engines and Groq touts deterministic latency, Tenstorrent’s bet on general-purpose programmability forces rivals to rebalance peak performance against model agility. Within 18 months, if its open-stack approach gains traction in mainstream AIGC frameworks, it could catalyze decentralized inference infrastructure—particularly in video generation, where hardware abstraction may finally outweigh transistor density as the key competitive lever.