Talos 2
A programmable NPU for 180 nm.
- Process target
- GF180 · 180 nm
- Weights
- Ternary: −1, 0, +1
- Activations
- Signed INT8
- Accumulation
- Signed INT32
- Development status
- Architecture, reference, and RTL work
Talos 2
Talos 2 is a programmable embedded inference accelerator targeting the GF180 180 nm manufacturing process. It is the second generation in the Talos neural chip family, moving toward tiled execution of trained dense layers using external model and working memory.
The current design uses ternary weights and INT8 activations, with a shared add/subtract/accumulate engine and INT32 partial sums. The intended host interface is SPI.
Fabrication is planned for December 2026.
Current work
The work includes a numerical contract, a software reference, and RTL components for arithmetic, buffering, scaling, and partial sums. Complete-engine integration, physical fit, timing, and end-to-end performance remain to be validated.
Spiking state updates are a possible later extension. Version 1 focuses on programmable dense inference.
Architecture
Schematic of the planned data path, based on the implementation plan. Status: in development.
