# Exchange costing and GeLU ownership, 2026-09-13 ## Model change Exchange estimates now separate payload bandwidth (4 bytes/cycle) from supervisor control issue work, taking their maximum rather than charging 160 cycles for every logical fragment. Low costing counts send controls and receive source/neutral controls, plus pointer resets inferred from relative allocation identities and offsets. Consecutive receive addresses retain the pointer; strided rows count resets arithmetically using the existing compressed geometry. Multicast payload is counted once at the sender. Ordinary transmitters have independent lanes; they no longer pessimistically share a lane with their neighbor. TX/RX controls on the same tile share supervisor issue capacity. The mid estimate lacks concrete receive addresses and uses three receive controls per estimated fragment. It retains the existing footprint calculation; changing cycle pricing does not remove row-storage screening. The local pack-versus-word-exchange decision uses the same payload/control price. This changes some concrete copy realizations even when a saved mid recipe is fixed. Low phase transport comparison also no longer walks geometry twice to account for the incorrect shared-TX assumption. This is a resource-work estimate, **not a schedule or a guaranteed bound**. Transfer order can change pointer continuation, paired mode depends on placement, and route latency, bank conflicts and dependency chains are not predicted. There is no new scheduling search, retry layer, or global correction multiplier. ## Calibration Artifacts are under `artifacts/exchange-cost-20260913/`: benchmark scripts and JSON, build logs, `calibration.csv`, `calibrate.py`, and `calibration.svg`. 48 logged phases from the saved BS1 recipe and 47 from the GeLU alternative were matched to their final scheduled horizons. Many phases are identical between these builds; these are not 95 independent workloads. Scheduled values below include the model's 600-cycle phase overhead, not hardware measurements. The comparison with the old fragment penalty uses the same new geometry and independent TX lanes, isolating the penalty rather than reproducing the entire old compiler. | Exchange | New estimate | Scheduled | Old fragment penalty | |---|---:|---:|---:| | Bias preparation | 2,752 | 3,331 | 39,040 | | Original GeLU → down projection | 10,360 | 10,769 | 21,440 | | Bias-owned GeLU → down projection | 10,360 | 10,765 | 39,040 | The two down-projection inputs both have a maximum 39,040-byte endpoint payload, but 134 versus 244 fragments. Their schedules differ by four cycles, not the 17,600 cycles implied by the old fragment charge. For the saved recipe, median estimate/scheduled is 0.827; median multiplicative error improves from 1.70× to 1.21×. The new estimates range from 0.40× to 0.98× actual scheduled duration. In particular, the original GeLU redistribution is still underestimated (1,676 versus 4,178 cycles). Conflict/ordering prediction remains an opportunity; these numbers do not justify hard feasibility bounds. Release/native-ISA runs with 24 Rayon threads: | Recipe | Mid costing | Low expansion including cost | Warm low re-cost | Footprint | |---|---:|---:|---:|---:| | BS1 original | 18.4 ms | 8.54 s | 1.30 s | 1.64 s | | BS1 bias-owned GeLU | 17.1 ms | 8.42 s | 1.29 s | 1.60 s | | BS2 saved | 18.5 ms | 11.78 s | 1.71 s | 2.25 s | Prior warm low re-cost was about 1.2 s / 1.6 s for BS1 / BS2. Some tests ran concurrently with these measurements and copy realizations changed, so this is an iteration-cost check, not a clean microbenchmark of estimator overhead. ## GeLU selection and hardware The saved extended BS1 recipe puts bias output on 729 row groups × 2 channel groups (1,458 tiles), then redistributes to 1,472 flat GeLU owners. Opening boundary 360 and reselecting GeLU retains bias ownership and removes that redistribution. Existing proposal generation already produces this alternative. Previously the coarse cost rejected it by 281,637 cycles. Now it predicts a 67,095-cycle improvement (13,118,410 → 13,051,315); unscheduled low costing predicts 77,166 cycles (10,456,164 → 10,378,998). The local-search trace confirms that this proposal is unvisited and passes the coarse gate. A limited shortlist can still prioritize other proposals ahead of it. The eight-step resumed search finished at 71 attempts in 396 seconds and accepted a different layout change (operation 12, boundary 350), reducing its placed estimate from 11,581,131 to 11,559,423 cycles. It did not select the GeLU ownership change within that budget. The GeLU hardware experiment below loads that alternative explicitly; it is not a claim that the short search found it. The search result was compiled but was not separately hardware-tested. Both recipes were built with the new compiler and run once on hardware: 27 layers, BS1, fused QKV, FP8 weights with scale -4, B1024 exchange ordering, FP32 reference, profiling of the first repeated layer. Profile times use the renderer's cropped origin. This isolates the ownership change from the earlier reduction bypass and from the new copy realization choices. | Measurement | Original GeLU ownership | Bias-owned GeLU | |---|---:|---:| | Cropped full-model cycles | 11,337,126 | 11,216,910 | | Cropped full-model time | 7.558084 ms | 7.477940 ms | | GeLU phase span | 14,328 | 11,928 | | Mean tile kernel cycles | 11,486.5 | 11,388 | | Maximum tile kernel cycles | 12,360 | 11,388 | | Kernel estimated useful utilization | 55.7% | 56.7% | | Device estimated useful utilization during GeLU | 44.6% | 53.6% | | FP32-reference cosine | 0.994264800 | 0.994264800 | This is a **1.06% full-model improvement**. The new GeLU kernel time is uniform across all 1,458 participating tiles; the remaining 540-cycle phase spread is start skew. Its estimate is 11,382 cycles, six below measurement. The old flat ownership includes irregular work lengths and slower scalar-pair tails. The arithmetic kernel itself is unchanged. Rendered profiles: `base-profile/model.html` and `gelu/model.html` in the artifact directory; corresponding packages, raw profiles, reference data and exact memory profiles are alongside them. The earlier historical extended-search profile was 11,499,450 cycles, but it also predates the reduction bypass and this cost-model change, so its 2.46% difference must not all be attributed to GeLU ownership. Bias–GeLU fusion is considered with the compatible ownership, but still rejected: 14,034 estimated cycles separately versus 15,726 fused. `GELU_WITH_BIAS` loads and adds the bias twice: once before the clamped polynomial and again to recover unclamped x+b for the final multiplication. It also reloads polynomial constants and performs per-row setup. Fusion avoids an add pass but increases inner-loop work. This rejection has not been independently benchmarked with a forced fused MLP kernel. Improving that kernel is a distinct task from making ownership compatible. GeLU itself now occupies about 2.7% of total execution cycles, so its remaining non-arithmetic overhead is not a large whole-model bottleneck. ## Validation The full codegen test run passed 281 tests and found one obsolete cost assertion: it expected 256 fragmented bytes to cost more than an 8-KiB contiguous transfer. That test now compares equal 8-KiB payloads. All 23 estimator tests then passed, including the corrected case and a new payload/control, pointer-continuation, and independent-neighbor test. Existing randomized geometry checks compare compressed receive-control counts against enumerated byte spans. Both full-model hardware/reference checks passed. No numerical kernels changed. Follow-up: [MIX bias–GeLU optimization](BIAS_GELU_MIX_2026_09_13.md) makes the fused kernel profitable and validates it on hardware. The rejection discussed above describes the older kernel.