Systolic Array vs. FEATHER
See how shape, mapping, and layout affect the work that reaches each PE.
Follow the same operands through a 16×16 systolic array and a 16×16 FEATHER NEST. Start with 7×12×32 (M×K×N), then follow a connected chain whose output becomes the next layer’s input. The seven-row chain also shows a visible producer–consumer layout change.
Same GEMM, different spatial work
Output-stationary
Adaptive spatial mapping
Same inputs. Same mathematical result.
Inspect one output from this layer. Both mappings evaluate the same deterministic operands; this is integer-valued mathematical GEMM, not bit-exact FP16 RTL execution.
Performance comparison by workload
Cycles are modeled logical cycles. PE utilization is useful MACs ÷ (256 × modeled cycles), including fill/drain.
| Workload shape M×K×N | Systolic array (SA) | FEATHER |
|---|
Mapping-dependent gains are not universal. Try the regular control and a weight-stationary or best modeled systolic baseline. Shape alone does not establish that FEATHER outperforms every systolic dataflow.
Layout switching is a separate question
A producer’s output order may differ from the next mapping’s input order. Watch the same output coordinates change bank/slot assignment below. A matching layout has zero additional repack traffic. Baseline assumptions and any copy cost are explicit, not hidden inside MAC utilization.
Each element keeps its ID and color in both buffers and during transfer. Select an element or click a buffer cell to follow its route.