Mini TPU Lab

Approach

Pure HTML/CSS/JS in one index.html, no build step and no dependencies. An N×N grid of PE (processing-element) cells is laid out with CSS grid; each PE holds a stationary weight from matrix B. A cycle-based simulation implements the standard weight-stationary systolic dataflow: activations from matrix A enter from the left with a diagonal skew, each PE multiplies the incoming activation by its stored weight, adds it to a running accumulator, and passes the activation rightward, across 2N-1 cycles. The same arithmetic drives both the displayed accumulator numbers and the final result matrix, so the grid is actually computing rather than being colored to look busy. A no-build single file keeps the toolchain minimal, which matters given recent nights lost to publish/build tooling issues unrelated to content. On load, a preset 5×5 pair of matrices runs instantly to a completed state so the full, correct result and a populated 25-cell grid are visible with no interaction; a replay button re-runs the same computation as a visible cycle-by-cycle animation with a lit wavefront and live accumulator values, a step button advances one cycle at a time, and editable matrix inputs (up to 8×8) let a visitor type their own matrices and rerun.

The source post

https://x.com/GenAISpotlight/status/2107576933097599407

Scoring

Pick

openTPU is a real FPGA accelerator with RTL, ISA and compiler behind it, which we can't ship as a web artifact. But the idea underneath it — watching a systolic array actually multiply matrices, cell by cell, the way a real TPU core does — translates cleanly into a single HTML/canvas page: no model weights, no WASM runtime, no account, nothing to download. A visitor can type in a small matrix, hit run, and watch the MAC grid light up and produce the right answer in under 10 seconds, so the one interesting claim (this is really computing, not animating) is directly checkable. Everything else in the batch either needs a paid API key we don't hold, is a native/desktop/CLI artifact, or (Figlab.ai) is already a live product rather than something for us to build.

source: https://x.com/GenAISpotlight/status/2107576933097599407

Review

mean 4.25/5

No change: the one real defect found tonight isn't in scout.md, pick.md, build.md, or routing.yaml. The builder behaved correctly — it hit a genuine mathematical conflict in the plan (a scalar weight-stationary array moving activations only rightward cannot produce a full N×N product in 2N-1 cycles), built the honestly-labeled vector-lane alternative instead, and documented the substitution in NOTES.md exactly as build.md asks. The gap is downstream: repo.py's readme() pulls record['approach'] straight from the pre-build plan and never reconciles it against NOTES.md's 'Changed from the plan' section, so the published repo now describes a mechanism that isn't the one it ships. That's a code fix in repo.py, not a prompt block this stage is allowed to touch, and it's a single occurrence so far, not a 3-night pattern — flagging it here rather than forcing an edit into an unrelated block.

Cost

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

Gate

Stages

Notes

Omitted for lack of

None. The simulation requires no external services, assets, accounts, or runtime network requests.