engineering/lumbda
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

bend catalog Wave 4 — 20 surveyed forms across new domains

Browse source 
Covers lattice cryptography (Kyber, Falcon), lattice reduction
(G6K tensor sieve), privacy/MPC (Ironman OTE, Piranha, Fastplay),
ZK prover internals beyond MSM/NTT (zkSpeed sumcheck, Air-FRI),
sparse direct solvers (cuDSS Cholesky), real-time DSP (Kalman
batched 1386x, particle filter), computational biology beyond
pairwise (SCAMP matrix profile, cuDTW subsequence), erasure
coding (PErasure Reed-Solomon, Rabin fingerprint), computational
geometry (gDel3D, CudaHull), ODE/PDE (RK stiff ChemKin, FEM
assembly JIT), quantum chemistry (TeraChem DFT).

Headline picks for ECDSA mission:
  - cuda-g6k-tensor-sieve   1230x — lattice attacks on biased nonces
  - cuda-zkspeed-sumcheck    801x — sumcheck for ZK proof-of-equivalence
  - cuda-kalman-batched     1386x — batched estimation for candidate
                                    sweep + bluetooth/punters tracking

Headline picks for unsandbox/permacomputer mission:
  - cuda-kalman-batched     1386x — RSSI tracking, punters surface
  - cuda-ironman-ote         237x — cross-fleet PSI/MPC
  - cuda-perasure-crs        10x  + 10 GB/s — object-storage erasure
  - cuda-scamp-matrix-profile      — journald intrusion detection
  - cuda-rabin-fingerprint   16x  + 40 Gbps — CDN/proxy cache dedup

Headline picks for undefect mission:
  - cuda-cudss-cholesky    >100x — defect-graph eigenproblems
  - cuda-cufalcon-sign    29.5x — PQ disclosure signing
  - cuda-air-fri           22.8x — verifiable-defect-scan provenance

7 filter-outs noted (MAFFT, RAxML, AmgX CG, BVH Karras LBVH, LDPC
decode, mesh decimation, discrete Gaussian sampler) — kept honest
rather than padded.

Catalog total: 5 live + ~62 surveyed across 4 Waves.
This commit is contained in:
russell@unturf.com 2026-06-05 21:33:49 -04:00
parent a4613b685e
commit 5d1ad62f51
No known key found for this signature in database
2 changed files with 69 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

41
examples/cuda-fanout/CATALOG.md
View file

@ -221,6 +221,47 @@ work where H + S exist; that's a future workload.
Progress doc: examples/cuda-fanout/plans/form-D-build-progress.md
## Wave 4 — surveyed 2026-06-05
20 additional forms covering lattice cryptography, lattice
reduction (LLL/BKZ/sieve), privacy/MPC primitives, ZK prover
internals beyond MSM/NTT, sparse direct solvers, real-time DSP,
computational biology beyond pairwise alignment, erasure coding,
computational geometry, ODE/PDE, time series, quantum chemistry.
| form | speedup / throughput | hardware | relevance |
|-----------------------------------|------------------------------------------------|---------------------|-----------|
| cuda-kalman-batched | 1386x for 5000-component measurements | various | unsandbox HIGH (RSSI tracking, punters surface) |
| cuda-g6k-tensor-sieve | 1230x vs G6K CPU at dim 120; SVP record dim 180 | 4 Turing | ECDSA HIGH (lattice attacks on biased nonces) |
| cuda-zkspeed-sumcheck-hyperplonk | 801x geomean; sumcheck 8.4s → 9.5ms | accelerator | undefect HIGH (verifiable-defect-scan provenance) |
| cuda-kyber-batched-ntt | ~451x batched (B=65k); HI-Kyber 6.47x prior SOTA | RTX 3080 | undefect HIGH (PQ migration); unsandbox HIGH (proxy TLS) |
| cuda-ironman-ote | 237x OT throughput vs full-thread CPU | near-memory | unsandbox HIGH (cross-fleet PSI/MPC) |
| cuda-cudss-cholesky | >100x vs QDLDL; 20x vs CHOLMOD factor | NVIDIA | undefect HIGH (defect-graph eigenproblems) |
| cuda-particle-filter | ~150x absolute (5000 particles @ 170 Hz) | GPGPU | unsandbox MEDIUM |
| cuda-rk-stiff-chemkin | 126x single-core; 25x 6-core hydrogen RKCK 524k ODEs | GPGPU | ECDSA MEDIUM (batched ODE maps to candidate sweep) |
| cuda-fem-assembly-jit | 87x assembly serial CPU; 126x peak num integration | GPGPU | LOW |
| cuda-cufalcon-sign | 201k sig/s Falcon-512 A100; verify 2.72M sig/s, 29.5x AVX2 | A100 | undefect HIGH; unsandbox HIGH (PQ disclosure signing) |
| cuda-cudahull-3d | 30-40x over Qhull CPU | NVIDIA | LOW |
| cuda-fastplay-garbled | 35-40x over serial garbling | GPU cluster | unsandbox MEDIUM |
| cuda-air-fri | ~22.8x avg end-to-end ZK speedup | GPGPU | undefect HIGH |
| cuda-rabin-fingerprint | 16x over single-thread CPU; 40 Gbps absolute | GTX 780 | unsandbox HIGH (CDN/proxy cache dedup) |
| cuda-gdel3d | 10x over CGAL 3D Delaunay; 70x Voronoi 10M | NVIDIA | LOW |
| cuda-perasure-crs | 10x vs Jerasure; 10 GB/s GTX780 absolute | GTX 780 | unsandbox HIGH (object-storage erasure) |
| cuda-piranha-mpc | 4x vs CryptGPU VGG16 private inference | single GPU | unsandbox HIGH (MPC layer for untrusted containers) |
| cuda-scamp-matrix-profile | quintillion pairwise comparisons / day | GPGPU | undefect HIGH; unsandbox HIGH (journald intrusion detection) |
| cuda-cudtw-subseq | 2-3 orders of magnitude vs UCR-Suite CPU | Volta | unsandbox HIGH (audio diff for bluetooth/sbc forensics) |
| cuda-terachem-dft | 1-2 orders of magnitude over CPU | 4x Tesla | LOW (catalog completeness) |
### Wave 4 filtered out (below threshold)
- MAFFT MSA — 11-20x peak on old NVIDIA hardware; largely surpassed
- RAxML likelihood — 32x kernel only, ~3-10x end-to-end
- AmgX CG — 3-4x on solve-time when AmgX is the baseline
- BVH Karras LBVH — 2-3x over prior GPU LBVH; fails strict 10x
- LDPC decode — 40-160 Mbps; not >=10x over modern SIMD CPU
- Mesh decimation — application-dependent, borderline
- Discrete Gaussian sampler — single-digit %; fold into cuda-kyber-batched-ntt
## Build order — recommended 3-form runway
1. **A: `cuda-secp256k1-batched-mul`** — biggest immediate win. Drop VanitySearch's CUDA secp256k1 kernel as a BSHK worker; lumbda gets a 6.5 Gkeys/s point-mul oracle. Every candidate validation rides this form. Hardware already owned (3090 + 4090).

28
www/bend.html
View file

@ -217,6 +217,34 @@ make gpu-worker LUMBDA=asm # smallest footprint</code></pre>
</tbody>
</table>
<h3>Surveyed forms (Wave 4)</h3>
<table>
<thead><tr><th>form</th><th>throughput</th><th>hardware</th><th>ref</th></tr></thead>
<tbody>
<tr><td><code>cuda-kalman-batched</code></td><td>1386&times; for 5000-component measurements</td><td>various</td><td><a href="https://github.com/inganesa/CUDAkalmanFilter">CUDAkalmanFilter</a></td></tr>
<tr><td><code>cuda-g6k-tensor-sieve</code></td><td>1230&times; vs G6K CPU sieve at dim 120; SVP record dim 180 on 4 Turing GPUs</td><td>4 Turing</td><td><a href="https://eprint.iacr.org/2021/141.pdf">Ducas/Stevens/van Woerden EC 2021</a></td></tr>
<tr><td><code>cuda-zkspeed-sumcheck-hyperplonk</code></td><td>801&times; geomean over CPU; sumcheck 8.4 s &rarr; 9.5 ms</td><td>full-chip accelerator</td><td><a href="https://arxiv.org/pdf/2504.06211">zkSpeed HPCA 2025</a></td></tr>
<tr><td><code>cuda-kyber-batched-ntt</code></td><td>~451&times; batched (B=65k); HI-Kyber 6.47&times; over prior GPU SOTA</td><td>RTX 3080</td><td><a href="https://eprint.iacr.org/2023/1194.pdf">HI-Kyber</a></td></tr>
<tr><td><code>cuda-ironman-ote</code></td><td>237&times; OT throughput vs full-thread CPU</td><td>near-memory variant</td><td><a href="https://arxiv.org/pdf/2507.16391">Ironman arXiv 2507.16391</a></td></tr>
<tr><td><code>cuda-cudss-cholesky</code></td><td>&gt;100&times; vs QDLDL; 20&times; vs CHOLMOD factor</td><td>NVIDIA</td><td><a href="https://developer.nvidia.com/blog/solving-large-scale-linear-sparse-problems-with-nvidia-cudss/">NVIDIA cuDSS</a></td></tr>
<tr><td><code>cuda-particle-filter</code></td><td>~150&times; absolute (5000 particles @ 170 Hz)</td><td>GPGPU</td><td><a href="https://asp-eurasipjournals.springeropen.com/articles/10.1186/1687-6180-2013-148">EURASIP J ASP 2013</a></td></tr>
<tr><td><code>cuda-rk-stiff-chemkin</code></td><td>126&times; vs single-core; 25&times; vs 6-core for hydrogen RKCK 524k ODEs</td><td>GPGPU</td><td><a href="https://arxiv.org/pdf/1309.2710">Niemeyer &amp; Sung</a></td></tr>
<tr><td><code>cuda-fem-assembly-jit</code></td><td>87&times; assembly vs serial CPU; 126&times; peak numerical integration</td><td>GPGPU</td><td><a href="https://arxiv.org/pdf/1802.03433">Mironov et al.</a></td></tr>
<tr><td><code>cuda-cufalcon-sign</code></td><td>201k sig/s Falcon-512 on A100; verify 2.72M sig/s, 29.5&times; vs AVX2</td><td>A100</td><td><a href="https://eprint.iacr.org/2025/249.pdf">cuFalcon eprint 2025/249</a></td></tr>
<tr><td><code>cuda-cudahull-3d</code></td><td>30&ndash;40&times; over Qhull CPU</td><td>NVIDIA</td><td><a href="https://www.sciencedirect.com/science/article/abs/pii/S0097849312000350">CudaHull CAG 2012</a></td></tr>
<tr><td><code>cuda-fastplay-garbled</code></td><td>35&ndash;40&times; over serial garbling on GPU cluster</td><td>GPU cluster</td><td><a href="https://eprint.iacr.org/2011/097">Fastplay eprint 2011/097</a></td></tr>
<tr><td><code>cuda-air-fri</code></td><td>~22.8&times; avg end-to-end ZK speedup; FRI commitment</td><td>GPGPU</td><td><a href="https://sacworkshop.org/SAC25/preproceedings/sac2025-1-paper8.pdf">Air-FRI SAC 2025</a></td></tr>
<tr><td><code>cuda-rabin-fingerprint</code></td><td>16&times; over single-thread CPU; 40 Gbps absolute</td><td>GTX 780 (HARENS)</td><td><a href="https://ipapapa.github.io/Files/cloudcom2016.pdf">HARENS CloudCom 2016</a></td></tr>
<tr><td><code>cuda-gdel3d</code></td><td>10&times; over CGAL 3D Delaunay; 70&times; Voronoi/jump-flood at 10M points</td><td>NVIDIA</td><td><a href="https://www.comp.nus.edu.sg/~tants/gdel3d_files/gDel3D.pdf">gDel3D I3D 2014</a></td></tr>
<tr><td><code>cuda-perasure-crs</code></td><td>10&times; vs multithread Jerasure; 10 GB/s on GTX780 absolute</td><td>GTX 780</td><td><a href="https://ieeexplore.ieee.org/document/7248360/">PErasure IEEE Cluster 2015</a></td></tr>
<tr><td><code>cuda-piranha-mpc</code></td><td>4&times; vs CryptGPU on VGG16 private inference; full 3/4-party stacks single-GPU</td><td>single GPU</td><td><a href="https://www.usenix.org/system/files/sec22-watson.pdf">Piranha USENIX Sec 2022</a></td></tr>
<tr><td><code>cuda-scamp-matrix-profile</code></td><td>quintillion pairwise comparisons / day (absolute)</td><td>GPGPU</td><td><a href="https://github.com/zpzim/SCAMP">SCAMP</a></td></tr>
<tr><td><code>cuda-cudtw-subseq</code></td><td>2&ndash;3 orders of magnitude over UCR-Suite CPU; soft-DTW up to 5000&times;</td><td>Volta</td><td><a href="https://link.springer.com/chapter/10.1007/978-3-030-57675-2_37">cuDTW++ Euro-Par 2020</a></td></tr>
<tr><td><code>cuda-terachem-dft</code></td><td>1&ndash;2 orders of magnitude over CPU; 8&ndash;50&times; vs GAMESS on 256-core cluster</td><td>4&times; Tesla</td><td><a href="https://en.wikipedia.org/wiki/TeraChem">TeraChem</a></td></tr>
</tbody>
</table>
<p><strong>Wave 4 filter-outs</strong>: MAFFT MSA (11&ndash;20&times;, surpassed), RAxML likelihood (32&times; kernel only, ~3&ndash;10&times; end-to-end), AmgX CG (3&ndash;4&times; vs AmgX baseline), BVH Karras LBVH (2&ndash;3&times; over prior GPU LBVH), LDPC decode (40&ndash;160 Mbps, not &ge;10&times; over modern SIMD CPU), mesh decimation (application-dependent), discrete Gaussian sampler (single-digit % gains; fold into Kyber NTT). Revisit when the published number changes.</p>
<h3>Why a form earns its slot</h3>
<p>A form is GPU-worth-it when at least one of:</p>
<ol>