ATIS
Before a single digital GPU cycle fires, analog circuits can determine which tokens in a transformer's attention window actually matter. ATIS uses Field-Programmable Analog Array hardware to perform importance scoring at microwatt scale — orders of magnitude below the energy cost of equivalent digital computation. By pre-filtering the attention matrix in analog, only the tokens that survive the importance threshold enter the digital pipeline, collapsing inference energy costs while preserving model accuracy. This architecture is the foundation of the patent portfolio.
Key Contributions
- Analog Transformer Importance Scoring (ATIS)
- FPAA hardware for microwatt-scale pre-filtering
- Patent portfolio for novel hardware architectures
Explainers
What is FPAA?
Field-Programmable Analog Arrays are reconfigurable analog circuits — the analog equivalent of FPGAs. They can be rewired to implement different continuous-time computations, running multiply-accumulate operations at microwatt power draws that digital hardware cannot approach. ATIS configures them as approximate attention scorers that exploit analog's natural affinity for the matrix operations underlying transformer self-attention.
How does analog pre-filtering help?
Transformer attention is O(n squared) — every token attends to every other token. ATIS introduces an analog gating layer that scores token-pair relevance before the digital attention computation begins. Tokens that score below the importance threshold are masked out, reducing the effective sequence length that the GPU must process. The energy savings compound quadratically with the number of filtered tokens.
What does this mean for edge inference?
ATIS makes transformer inference viable on devices with strict power budgets — embedded systems, mobile hardware, IoT sensors. A microwatt analog frontend can gate which tokens reach a milliwatt digital backend, enabling large-context model inference in environments where running full attention would drain a battery in minutes.