Skip to content

artifact: add the-noise-ceiling - #173

Merged
blokzdev merged 2 commits into
mainfrom
claude/friendly-einstein-p5cq8l
Aug 1, 2026
Merged

artifact: add the-noise-ceiling#173
blokzdev merged 2 commits into
mainfrom
claude/friendly-einstein-p5cq8l

Conversation

@blokzdev

@blokzdev blokzdev commented Aug 1, 2026

Copy link
Copy Markdown
Owner

Artifact: The Noise Ceiling

An interactive simulation of the FHE noise budget — the mechanism The Bootstrapping Tax hinges on but never depicts. Paired with (and embedded in) the merged article the-bootstrapping-tax-fhes-fundamental-limit-on-private-ai-inference.

What it shows

Every FHE multiplication spends down a finite budget of "levels"; when noise reaches the decryption ceiling, the ciphertext must be bootstrapped — refreshed back down. The stage animates that noise sawtooth as you drag the circuit deeper, and the scheme toggle makes the article's central tradeoff concrete:

  • CKKS — large budget (L≈16), climbs for many operations, then pays ~1 min per refresh → few, expensive teeth.
  • TFHE — refreshes on every gate for ~1 ms → constant, cheap teeth (budget slider dims; L≡1).

The panel reports the live regime (leveled ✓ / bootstrapped ×N), the bootstrap count, and total bootstrapping time, plus measured per-model FHE-vs-MPC latency cards (BERT-Base 149 s / 112 GB · ResNet-50 164 s / 228 GB · BERT-Tiny · ResNet-20 · LogReg) that anchor the mechanism to real numbers. Tap a card to pin its detail (memory, MPC LAN/WAN, throughput gap).

Archetype

simulation (svg) — a mechanism unfolding over time; underrepresented against the catalog's chart/diagram/calculator majority.

Data source

content/artifacts/the-noise-ceiling/data.json, snapshotted from the article's cited primaries — arXiv:2604.00169 Beyond Latency and arXiv:2504.11604 SoK: Can FHE Support General AI Computation?dataAsOf 2026-06-23. No runtime fetches. The sawtooth is a first-order model of leveled evaluation + bootstrapping; the per-model latency/memory figures are measured, not model output (stated in the caption).

Verified

  • npm run validate + npm run check (astro types) + npm run build all green (CI parity).
  • Chromium (Playwright) load at 340 / 390 / 900 px: zero console errors, zero horizontal overflow; portrait / landscape / narrow render cleanly, CKKS sawtooth resets to level 2 after each bootstrap, TFHE shows dense full-height teeth, budget slider dims under TFHE.
  • Interaction paths: scheme toggle, depth & budget sliders (native keyboard), Evaluate sweep, Reset, clickable model cards.

Duplicate check

Re-checked all 138 prior artifacts, both open PRs, and last-24h merges. the-fhe-inference-wall (chart) covers the latency wall, not the noise/bootstrapping mechanism; the-noise-floor is an unrelated RL-sample-complexity calculator. Open PR #169 is prediction-market fragmentation. No commission-artifact issues open.

@vercel

vercel Bot commented Aug 1, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
blokz Ready Ready Preview Aug 1, 2026 9:22am

Simulation of the FHE noise budget + bootstrapping mechanism, paired with
the-bootstrapping-tax. Animated noise sawtooth under a decryption ceiling;
CKKS vs TFHE scheme toggle contrasts rare-but-slow vs constant-but-cheap
bootstrapping. Measured per-model FHE/MPC latencies from arXiv:2604.00169 and
arXiv:2504.11604 anchor the mechanism.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NtUu7oGdyHAwEecPY1cWz4
@blokzdev
blokzdev marked this pull request as ready for review August 1, 2026 09:21
@blokzdev
blokzdev enabled auto-merge (squash) August 1, 2026 09:21
@blokzdev
blokzdev merged commit b29bbc6 into main Aug 1, 2026
3 checks passed
@blokzdev
blokzdev deleted the claude/friendly-einstein-p5cq8l branch August 1, 2026 09:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants