PFN, Tenstorrent launch joint generative-AI processor system work

Tenstorrent brings a track record in chip design and an open-source instruction set: it develops processors based on RISC-V and already ships the Blackhole AI processor and the Ascalon RISC-V CPU. In Japan, it is also working with Rapidus on a 2nm logic edge AI accelerator.

PFN's side of the work draws on its own processor line. Its MN-Core is optimized for matrix arithmetic, and the company says it has taken the top spot in the supercomputer energy-efficiency ranking three times. The next-generation L1000, announced in November 2024, is designed around wide memory bandwidth to speed LLM inference, and PFN has said a workstation with L1000 can reach roughly 500 to 1,000 tokens per second on a 70B-parameter model. The L1000's planned availability has slipped from 2026 to 2027.

The high-memory-bandwidth AI processor field already has early movers in Cerebras and Groq. Cerebras works with OpenAI and is part of the inference base for ChatGPT's Ultrafast mode, while Nvidia has acquired a non-exclusive license to Groq's technology for $20 billion, with Groq's chips to be offered as part of the next-generation Vera Rubin AI supercomputer. Photonics-electronics convergence, which uses light rather than electricity where possible to move data with lower power consumption and higher capacity, is presented as key to computing efficiently across multiple nodes.