IEEE TCAS-I 2026 / INTERACTIVE RESEARCH COMPANION
Attention,
inside memory.
Follow a query from binary matches to a weighted answer. Then explore the hardware tradeoffs behind CAMformer.
Published operating point
Simulation-based attention-layer results
FROM BITS TO AN ANSWER
One bit can change who gets attention.
Click a query bit. Matching keys light up, selected scores become weights, and the output updates.
Stored keys
Match RetainedCONTEXTUALIZATION
A weighted answer
Selected values contribute in proportion to their attention weights.
Synthetic values; floating-point demo arithmetic. The hardware uses BF16 values and MACs. This is not a trained-language-model evaluation.
THE PHYSICAL INTUITION
A match becomes a contribution.
Choose a stored row and flip query bits. Inspect the comparison before following the summed response.
Query bits
Conceptual signal flow, not a transistor-level schematic or SPICE simulation.
ILLUSTRATIVE IDEAL RESPONSE
Normalized response = matching bits / 8. This illustrates charge-sharing similarity sensing; it does not predict absolute voltage, ADC error, or PVT behavior.
FOLLOW THE DATA
Three stages. One attention operation.
Step through the architecture, from query-key association to high-precision value aggregation.
Default stage service times
READING THE TIMING
Latency and throughput answer different questions.
Bars are exported from the release model at 1 GHz. Walkthrough animation is explanatory, not a cycle-exact event trace. Single-core MHA throughput includes 16 attention heads.
TABLE IV · VALIDATED SWEEPS
Find the tradeoff behind the default.
Compare the released configurations. Every point comes directly from the Python hardware model.
Throughput and efficiency
What changes?
Expected recall is a hypergeometric estimate under uniformly located global top-k entries. It is not task accuracy. Runtime top-k reuses fixed Top-32 hardware, so its modeled area stays constant.
EXPLORE THE EVIDENCE
The paper, with room to inspect.
Hover for exact values. Toggle traces, zoom, and export a figure or its underlying data.
Compare with the original author figure
BUILT ON REPRODUCIBLE RESEARCH
Every number has a source.
Published values, release-model outputs and illustrative interactions are labeled separately.