A visual tour of what happens when a 100-tree, depth-4 boosted decision tree stops being software and becomes an ultra-low-latency decision circuit on a Xilinx VU9P.
The trigger problem is about arranging nanoseconds and hardware blocks, not maximizing software throughput. This view shows where the BDT sits in the path from collision to accept/reject logic.
Threshold comparisons and branch routing map naturally to LUT fabric. The BDT spends logic on control and accumulation, while a dense MLP spends more of its budget on multiply-accumulate units and DSP blocks.
Hover any stage to inspect local timing and resource emphasis.
Each split becomes a comparator, each branch becomes signal routing, and each leaf becomes a small fixed-point score contribution. Step through the conversion and inspect the path logic.
No pointer chasing, no software walk. The tree is unrolled so traversal becomes combinational signal flow, then class scores are summed across the ensemble.
Hover comparators and leaves to reveal the fixed-point threshold or score value.
The paper’s useful operating region is not “full precision or bust.” This section shows a mock but paper-shaped view of class retention as bit width shifts from fragile to comfortable.
At very low precision, thresholds and leaf outputs drift enough to bend operating points. In the 11–15 bit region, the model still behaves much like the floating-point reference while staying inside a tolerable hardware budget.
Switch the overlay to view retention, threshold jitter, or resource savings.
The BDT does not win by raw classifier quality. It wins by landing in a different resource shape: more LUT-heavy, less DSP-hungry, and still fast enough to fit the trigger window.
The BDT burns more of the configurable logic fabric on decisions and reductions. The MLP concentrates more demand in multipliers and DSP blocks. Which is better depends on the board budget around the model, not on a single headline metric.
Toggle between model metrics and a frequency-depth pressure map.
Selected sources behind the visuals and episode framing.