🧠 Food for thought

Implications, context, and why it matters.

A veteran team is betting on a throwback chip design

• The founding group includes alumni from AMD and Tenstorrent. CEO Ljubisa Bajic previously founded Tenstorrent 1.
• The chip locks an AI model’s weights into a “mask ROM recall fabric.” It pairs that with an SRAM recall fabric using on-chip Static Random-Access Memory for computation plus changing data such as the KV (key-value) cache, a memory used by transformer models to speed up text generation 1.
• Each major model update needs a new chip spin. Taalas says porting its HC inference engine to a new model release means changing two metal layers, while training a model costs about 100x more than producing a customized HC chip in reasonable volumes 1.
• Taalas targets a two-month manufacturing cycle by building an almost complete chip, then customizing two metal layers at its foundry partner, TSMC (Taiwan Semiconductor Manufacturing Co.), on its manufacturing process 2.

Taalas’s success could turn AI models into physical products

• If the approach works, models could ship as physical chips rather than software accessed via an API.
• Buyers would give up easy model updates in exchange for lower costs plus higher efficiency on a stable, tested version 3.
• Large model developers could order custom accelerators that preserve a popular release for sale as a specialized appliance. Another major revision would still require a new chip spin 1.
• This puts pressure on cloud providers that rent general-purpose GPUs. Taalas says dedicated inference hardware could beat a “small GPU data center” on a single chip 3.