Live matmul arena · attention prototype roadmap · Bittensor SN74

An open arena for cheaper matrix multiplication.

C = A × B is the hot loop of nearly all numerical computing. CCO is where you submit a matmul strategy that computes it for less — lower latency, lower VRAM, lower time-complexity — without giving up accuracy, and have that claim measured the same way for everyone.

Whitelisted on Gittensor (Bittensor SN74). Merged strategies earn TAO through Gittensor's emission share for this repository. No Discord access or prior SN74 knowledge required — this repository is the whole story.
Deterministic, hardware-measured benchmarks — no LLM-as-a-judge.
strategy/examples/run_example.py
# Normal (exact) computing — O(N³), the frontier you must beat
- C = A @ B

# Smart (subspace) strategy — compress, compute, reconstruct
+ from strategy import subspace_matmul, Config
+ C = subspace_matmul(A, B, config=Config(transform="rsvd", rank_m=128))

assert rel_err(C, A @ B) < 1e-2   # admitted only if accurate AND cheaper
# O(N²M) vs O(N³) · scored on accuracy, latency, VRAM, FLOPs — together
Gittensor

CCO earns through Gittensor — Bittensor subnet 74.

Gittensor pays TAO for real, merged pull requests to whitelisted open-source repositories. CCO is one of them. You don't need Discord access or prior SN74 knowledge to take part — read the docs here, run the scorer on your own GPU, and open a PR through the same GitHub path as any open-source project. This repository stays the single source of truth for the code, the scorer, the docs, and every admitted improvement.

O(N³)→O(N²M)
Complexity a strategy must beat
4
Cost axes scored together — never traded off
SN74
Bittensor subnet — Gittensor
1 + prototype
Live matmul track, attention staged separately
The problem

Matrix multiply cost grows with the cube of size.

C = A × B costs 2N³ FLOPs computed exactly. The same matrix operations sit underneath attention layers in Transformer LLMs — which is why CCO's roadmap points toward an attention-shaped track — but today's score-bearing arena is the general, provider-agnostic case: exact matrix multiplication, measured honestly.

2 GFLOP
N = 1,000

Trivial today — the N³ term means this is not where the real cost lives.

3.5 TFLOP
N = 12,000 — CCO's reference regime

Full-rank data, deliberately the hard case — no structure to exploit for free.

4.2 PFLOP
N = 128,000 — out-of-core scale

Where the exact engine must tile across disk-backed memory — a cheaper algorithm stops being optional.

How it works

Submit a strategy. Get scored. Get paid for what genuinely improves.

No prize pool, no cycles, no committee — a deterministic scorer decides, and Gittensor's emission share pays for what it admits.

01

Submit a matmul strategy

Add a Transform to strategy/transforms.py and open a PR with your self-scored scorecard.

02

Deterministic scoring

python -m eval reruns your matmul strategy against the exact baseline on identical inputs — accuracy, latency, VRAM, FLOPs, together.

03

Merge on merit

Admitted only if it dominates the exact baseline on every cost axis while holding accuracy — no averaging a win on one axis against a loss on another.

04

Gittensor pays

SN74 validators see the merged PR and pay TAO from CCO's emission share — no separate application, no cycle to wait for.

The technology

Compress, compute, reconstruct.

The reference strategy projects A and B onto a shared M-dimensional subspace (M ≪ N), multiplies the small M×M cores, and reconstructs — O(N²M) instead of O(N³). The subspace basis is the pluggable part: a Transform you write and register.

  • Exact baseline, always. Every strategy is compared to real A @ B on identical random inputs — never a synthetic or cherry-picked case.
  • Honest by construction. The reference regime is full-rank data — no low-rank structure to exploit for free. Beating it means finding real structure.
  • One registration away. Subclass Transform, implement basis(), call register_transform() — that's the whole interface.
strategy/subspace.py
à = Qᵀ A Q    B̃ = Qᵀ B Q    C̃ = à B̃    C ≈ Q C̃ Qᵀ

Q is an (N, M) orthonormal basis — the only thing a contributed Transform supplies. Everything else in the pipeline is shared, tested infrastructure.

EXACT
O(N³)
SUBSPACE
O(N²M), M ≪ N
Deterministic evaluation

Scored by physics, not opinion.

Every submission passes through the same automated gate on identical inputs. No language model ever judges another strategy — ranking is pure measurement, reproducible by anyone with a GPU.

Tier 1 · Accuracy floor

Bounded Frobenius accuracy against the exact product. Fall below the floor and the score is zero — a fast wrong answer is not an improvement.

Tier 2 · Scaling proof

An empirical N^p fit from --sweep. A claimed sub-cubic strategy must show it in measured latency, not just in algebra.

Tier 3 · Dominance gate

Admitted only if latency, VRAM, and FLOPs all beat the exact baseline while accuracy holds. One regressing axis disqualifies it — no averaging a win against a loss.

Roadmap

From matmul, live today, to sub-quadratic attention.

Progress is gated by achievement, not a calendar. Matmul is the open score-bearing arena running right now. Attention is being staged separately as a prototype until the evaluator, bot, GPU validator, and dashboard are track-aware.

Phase 0Live now

Matmul arena

Approximate and exact-but-cheaper strategies for C = A × B, scored on full-rank, low-rank, and decaying-spectrum regimes. Open for contributions today.

subspaceN=12000open now
Phase 1Prototype now

Foundation — up to 32K tokens

Adapt the matmul-arena discipline to attention: exact attention baseline, synthetic Q/K/V inputs, hybrid local-window plus spectral-global prototypes, then RTX 5090 validation once the track is official.

attention/spectral mixingnot scored yet
Phase 2Roadmap

Log-linear — up to 128K tokens

Novel hybrid approaches required; spectral or low-rank kernels needed to pass the scaling gate. MSE < 0.005.

spectrallow-rank128K ctx
Phase 3Roadmap

Million-token — up to 1M tokens

Production-grade kernels with true sub-quadratic VRAM scaling and hardware-specific Triton/CUDA tuning. MSE < 0.001.

Triton/CUDA1M ctx
Phase 4Ongoing

Convergence — arbitrary context

Fine-grained hardware tuning and optional weight co-optimization (LoRA patches) to recover residual approximation error. MSE < 0.0005.

LoRA patcheshardware-tuned
Get involved

Two ways to take part today.

Contribute a matmul strategy

Add a Transform, self-score it, open a PR. Admitted matmul improvements are merged and — through Gittensor — paid.

  • One rule: every cost axis down, accuracy held
  • Reviewed and merged on the reproduced scorecard
Read CONTRIBUTING.md

Run the scorer

Clone the repo, install PyTorch, and reproduce the matmul reference scorecard on your own GPU — the same numbers a reviewer checks.

  • CUDA or Apple MPS, PyTorch-only
  • python -m eval --n 12000 --pairs 3
See the quickstart

The frontier is open.

Read the whitepaper to see exactly how scoring and rewards work, or jump straight to contributing a matmul strategy — the scorer decides, not us.