Search. Route. Learn.

K1 searches a 172-architecture INT8 GEMM design space, then uses final-route timing to decide what to test next.

The Kernellum Difference

Routed search
Use workload structure to propose candidates, then test the predictions against synthesis and final-route timing.
Physical feedback
Update the surrogate with routed measurements instead of treating pre-layout estimates as ground truth.
Held-out falsification
Freeze discovery and deployment cohorts, preregister criteria, and report only what survives unseen routes.

Measured Evidence

K1 routed 9/9 frozen architectures; mean predicted-vs-routed workload ranking was ρ = 0.944, with 0.583% mean regret for the analytically selected winner.
In the 172-architecture closed-loop study, active search improved all 12 Transformer GEMMs. Random improved none under the same four-candidate budget.
Across 96/96 ECP5 routes, operand-distribution topology produced a predictable timing effect. Held-out tax prediction reached 0.3039 ns MAE and 0.8921 correlation.
A two-candidate local portfolio beat broadcast-only in 36/36 unseen cases, with 23.56% mean latency improvement and 1.44% mean oracle regret.

Research Loop

Map the workload
Understand the workload signals that make an architecture efficient before committing to RTL.
Search architectures
Test how dataflow, memory, and topology choices change resource use and final-route timing.
Measure implementation
Measure the physical cost of each design and keep the evidence attached to the decision.
Close the loop
Update the search as implementation changes, so the loop learns what physical design actually rewards.