Most AI accelerators are black boxes. You can only run what the vendor has optimized. Tenstorrent Blackhole is different: everything is open source, from low-level kernels written for its custom Tensix ISA all the way up to the compiler and PyTorch/vLLM integration.
In this talk, we’ll show what that openness gets you. Blackhole runs agentic, generative and physical AI well, from 2T+ LLMs to video generation to robotics. New models can be ported and optimized within a day of release, using AI skills. And because the cores are fully programmable, you can go beyond AI to things like graphics, simulation and physics solvers.
You’ll leave knowing how the stack fits together and how to try it yourself.
