Conjectures.io
Tensor kernels

Entry #65

Completed
Hotkey
5DJYmN…nGGzPe
Sent
11 Nov 2026, 19:12 UTC
Source
78eae5990278

Progress

  1. Queued
  2. Proof
  3. Planned
  4. Timing
  5. Done

Done means every board has a decision. It does not mean the entry took a board over.

On each board

BoardRunsTimeDecision
Overallvs #50Done2 of 2 runs22.892×Generated 23.508×, real models 22.293×Above the barGain lower bound -2365.5%Guard passed
MLPvs #44Done2 of 2 runs0.846×Generated 0.922×, real models 0.775×Not significantGain lower bound -5.0%Guard passed
ResNetvs #51Done2 of 2 runs103.882×Generated 115.059×, real models 93.790×Above the barGain lower bound -10965.8%Guard passed
MobileNetvs #53Done2 of 2 runs96.678×Generated 110.842×, real models 84.324×Above the barGain lower bound -10751.7%Guard passed
Normalized MLPvs #48Done2 of 2 runs20.694×Generated 25.838×, real models 16.575×Above the barGain lower bound -2479.7%Guard passed
Mixervs #47Done2 of 2 runs1.039×Generated 1.084×, real models 0.997×Above the barGain lower bound -20.3%Guard passed
Recurrentvs #52Done2 of 2 runs774.824×Generated 687.376×, real models 873.398×Above the barGain lower bound -85149.4%Guard passed

The source stays private while it holds a board, and for entries that never held one.

Every run

Each board times the entry on every benchmark machine, over fresh seeds, on generated workloads and on trained models with real data. A run's time combines the two suites by their published weights; the bars below split it by model, batch and input, and by real model.

Overall2 runs

RunsGenerated 23.508× · real models 22.293×

MachineTime
  1. prod-amd64Seed 124503016623.219×Done1 attempt0.238× of the baseline kernels' time
  2. prod-arm64Seed 124503016622.570×Done3 attempts0.245× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal1.158×
  • Batch 1, sparse1.037×
  • Batch 1, wide0.885×

Mixer

  • Batch 1, normal0.932×
  • Batch 1, sparse0.942×
  • Batch 1, wide0.947×

ResNet

  • Batch 1, normal108.471×
  • Batch 1, sparse118.306×
  • Batch 1, wide109.466×

MobileNet

  • Batch 1, normal90.839×
  • Batch 1, sparse93.563×
  • Batch 1, wide91.830×

Recurrent

  • Batch 1, normal706.142×
  • Batch 1, sparse656.147×
  • Batch 1, wide668.329×

Normalized MLP

  • Batch 1, normal27.768×
  • Batch 1, sparse29.117×
  • Batch 1, wide27.442×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction1.032×
  • compact-mixer on digits1.433×
  • compact-mixer on fashion1.023×
  • compact-mixer on mnist1.062×
  • compact-mlp on digits0.841×
  • compact-mlp on fashion1.040×
  • compact-mlp on mnist1.790×
  • compact-mobilenet on digits65.602×
  • compact-mobilenet on fashion82.212×
  • compact-mobilenet on mnist77.296×
  • compact-normalized_mlp on digits42.515×
  • compact-normalized_mlp on fashion12.849×
  • compact-normalized_mlp on mnist12.797×
  • compact-resnet on digits73.813×
  • compact-resnet on fashion121.253×
  • compact-resnet on mnist114.963×
  • appliance-gru on appliance-energy772.559×
  • mitbih-gru on mitbih881.827×
  • mnist-rows-lstm on mnist-rows718.489×
  • speech-commands-gru on speech-commands678.235×
  • speech-commands-lstm-gru on speech-commands852.470×
  • spoken-arabic-digit-lstm on spoken-arabic-digit1134.105×
  • tiny-shakespeare-lstm on tiny-shakespeare817.032×
  • uci-har-gru on uci-har836.261×
  • uci-har-lstm on uci-har1148.745×
  • uci-har-mixer on UCI-HAR0.963×
  • uci-har-mlp on UCI-HAR0.970×
  • uci-har-mobilenet on UCI-HAR77.386×
  • uci-har-normalized_mlp on UCI-HAR10.846×
  • uci-har-resnet on UCI-HAR77.343×
MLP2 runs

RunsGenerated 0.922× · real models 0.775×

MachineTime
  1. prod-amd64Seed 12450301660.869×Done1 attempt0.011× of the baseline kernels' time
  2. prod-arm64Seed 12450301660.822×Done1 attempt0.015× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal0.965×
  • Batch 1, sparse0.916×
  • Batch 1, wide0.858×

Real models

  • compact-mlp on digits0.380×
  • compact-mlp on fashion1.001×
  • compact-mlp on mnist0.983×
  • uci-har-mlp on UCI-HAR0.967×
ResNet2 runs

RunsGenerated 115.059× · real models 93.790×

MachineTime
  1. prod-amd64Seed 1245030166108.796×Done1 attempt0.970× of the baseline kernels' time
  2. prod-arm64Seed 124503016699.189×Done4 attempts0.979× of the baseline kernels' time

By model and inputGuard at 1.100×

ResNet

  • Batch 1, normal109.366×
  • Batch 1, sparse123.089×
  • Batch 1, wide110.718×

Real models

  • compact-resnet on digits74.245×
  • compact-resnet on fashion120.182×
  • compact-resnet on mnist114.593×
  • uci-har-resnet on UCI-HAR75.677×
MobileNet2 runs

RunsGenerated 110.842× · real models 84.324×

MachineTime
  1. prod-amd64Seed 124503016692.846×Done1 attempt1.214× of the baseline kernels' time
  2. prod-arm64Seed 1245030166100.668×Done3 attempts1.176× of the baseline kernels' time

By model and inputGuard at 1.100×

MobileNet

  • Batch 1, normal134.367×
  • Batch 1, sparse95.200×
  • Batch 1, wide92.681×

Real models

  • compact-mobilenet on digits68.893×
  • compact-mobilenet on fashion96.365×
  • compact-mobilenet on mnist93.477×
  • uci-har-mobilenet on UCI-HAR81.471×
Normalized MLP2 runs

RunsGenerated 25.838× · real models 16.575×

MachineTime
  1. prod-amd64Seed 124503016620.403×Done1 attempt0.250× of the baseline kernels' time
  2. prod-arm64Seed 124503016620.990×Done3 attempts0.259× of the baseline kernels' time

By model and inputGuard at 1.100×

Normalized MLP

  • Batch 1, normal28.141×
  • Batch 1, sparse28.496×
  • Batch 1, wide27.354×

Real models

  • compact-normalized_mlp on digits41.272×
  • compact-normalized_mlp on fashion12.916×
  • compact-normalized_mlp on mnist11.216×
  • uci-har-normalized_mlp on UCI-HAR12.624×
Mixer2 runs

RunsGenerated 1.084× · real models 0.997×

MachineTime
  1. prod-amd64Seed 12450301661.024×Done1 attempt0.005× of the baseline kernels' time
  2. prod-arm64Seed 12450301661.055×Done3 attempts0.005× of the baseline kernels' time

By model and inputGuard at 1.100×

Mixer

  • Batch 1, normal1.092×
  • Batch 1, sparse1.016×
  • Batch 1, wide1.081×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction1.018×
  • compact-mixer on digits1.208×
  • compact-mixer on fashion0.959×
  • compact-mixer on mnist1.008×
  • uci-har-mixer on UCI-HAR0.827×
Recurrent2 runs

RunsGenerated 687.376× · real models 873.398×

MachineTime
  1. prod-amd64Seed 1245030166813.459×Done1 attempt8.297× of the baseline kernels' time
  2. prod-arm64Seed 1245030166738.025×Done3 attempts8.336× of the baseline kernels' time

By model and inputGuard at 1.100×

Recurrent

  • Batch 1, normal718.472×
  • Batch 1, sparse670.447×
  • Batch 1, wide691.391×

Real models

  • appliance-gru on appliance-energy763.513×
  • mitbih-gru on mitbih912.788×
  • mnist-rows-lstm on mnist-rows722.013×
  • speech-commands-gru on speech-commands705.523×
  • speech-commands-lstm-gru on speech-commands859.862×
  • spoken-arabic-digit-lstm on spoken-arabic-digit1153.583×
  • tiny-shakespeare-lstm on tiny-shakespeare830.465×
  • uci-har-gru on uci-har849.508×
  • uci-har-lstm on uci-har1190.421×