Skip to content

Task 12: 2D-grid Dally model extension to sparse-parity

Priority: HIGH Status: OPEN Agent: unassigned Source: Telegram chat-yaroslav, 2026-05-08T00:56:38Z — "Looking into extending sparse-parity challenge to use the 2D grid model. Might need adding instructions to https://github.com/cybertronai/simplified-dally-model , trying to see how few I can get away [with]"

Context

The current sparse-parity benchmark + ByteDMD scoring assumes a 1D memory hierarchy (registers → L1 → L2 → HBM with hardcoded per-level pJ costs from Bill Dally's numbers). Yaroslav is now extending to a 2D grid model (cybertronai/simplified-dally-model) where memory locality is two-dimensional — closer to actual physical layout of arithmetic units on a die — and exploring how few primitive instructions are needed to capture the relevant cost dynamics.

This affects the sparse-parity challenge in two ways: 1. The scoring metric changes — submissions need to be re-scored against the 2D-grid model, not just the 1D Dally-numbers heuristic. 2. The instruction surface for tracked code may shrink — a smaller primitive set means fewer ways for an algorithm to "leak" untracked operations.

Relevance to SutroYaro: SutroYaro is the harness + scoring + leaderboard for sparse-parity. If the underlying physical model changes, the entire scoring pipeline (src/bytedmd/, src/sparse_parity/eval/, the eval-environment) needs to follow.

References: - cybertronai/simplified-dally-model (the 2D grid + reduced-instruction-set repo Yaroslav is iterating on) - ByteDMD spec/reference: https://github.com/cybertronai/ByteDMD - 1D Dally numbers (current model): register 5 pJ, L1 (64KB) 20 pJ, L2 (256KB) 100 pJ, HBM 640 pJ per float access

Tasks

Phase 1 — Read + summarize the 2D-grid model

  • Read cybertronai/simplified-dally-model README + any current iteration notes
  • Document: what's the 2D-grid coordinate system? what costs does it compute? what's the minimum instruction set Yaroslav has converged on?
  • Compare against the current 1D Dally cost model in src/sparse_parity/cache_tracker.py
  • Write to docs/research/2d-grid-model-summary.md

Phase 2 — Identify which sparse-parity solvers re-score most differently

The current top solvers (RESULTS) are KM-min, GF(2) Gaussian elimination, KM influence estimation, SMT backtracking, SGD baseline. Each has a distinct memory-access pattern: - KM-min: ~20 reads / 3,578 DMC — small, cache-friendly - GF(2): ~420 reads / ~203K DMC — k-independent but heavier - SGD: 8,504 reads / 1,278,460 DMC — many activations, gradient-add re-fetches

Under 2D-grid scoring, locality matters more (or differently). Hypothesis: the gap between KM-min and SGD widens further; GF(2)'s position shifts depending on whether matrix-fused operations register as nearby in the 2D layout.

  • Re-score top 5 solvers under the 2D-grid model
  • Diff against current 1D ranking
  • Write to docs/findings/2d-grid-rescoring.md

Phase 3 — Submission-pipeline plumbing

  • If the 2D-grid model is going to become the primary metric, draft the change to the submission pipeline at cybertronai/sparse-parity-challenge (issue, then PR)
  • Backwards-compat: keep 1D ByteDMD as a fallback metric for now; add a new --metric=2d-grid flag to the eval harness
  • Update docs/research/eval-environment.md to document both

Acceptance

  • A docs/research/2d-grid-model-summary.md that any new agent can read to understand what changed
  • A docs/findings/2d-grid-rescoring.md showing top-N solvers under both metrics side-by-side
  • A draft plan for the eval-harness change (PR or issue link)

Dependencies

  • cybertronai/simplified-dally-model must reach a stable instruction set first (Yaroslav is still iterating per Telegram)
  • This task should not start Phase 3 until Yaroslav explicitly signals "2D grid is the new primary"