Conjectures.io
Tensor kernels

Entry #5

Completed
Hotkey
5Ct5xu…fFPn5t
Sent
10 Nov 2026, 18:42 UTC
Source
69acfc349dec

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 #8Done2 of 2 runs0.834×Generated 0.724×, real models 0.961×QualifiesGain lower bound 10.1%Guard passed
MLPvs #1Done2 of 2 runs1.060×Generated 1.033×, real models 1.087×Above the barGain lower bound -35.3%Guard passed
ResNetvs #9Done2 of 2 runs0.994×Generated 0.996×, real models 0.992×Above the barGain lower bound 0.3%Guard passed
MobileNetvs the shipped referenceDone2 of 2 runs0.937×Generated 0.623×, real models 1.411×Not significantGain lower bound -11.9%Guard passed
Normalized MLPvs #6Done2 of 2 runs1.011×Generated 1.012×, real models 1.011×Above the barGain lower bound -15.3%Guard passed
Mixervs #7Done2 of 2 runs1.047×Generated 1.068×, real models 1.027×Above the barGain lower bound -21.4%Guard passed
Recurrentvs #10Done2 of 2 runs16.207×Generated 15.250×, real models 17.224×Above the barGain lower bound -1559.6%Guard passed

Download Kernel.lean

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 0.724× · real models 0.961×

MachineTime
  1. prod-amd64Seed 11563760320.807×Done1 attempt0.103× of the baseline kernels' time
  2. prod-arm64Seed 11563760320.862×Done1 attempt0.115× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal1.124×
  • Batch 1, sparse1.006×
  • Batch 1, wide1.237×

Mixer

  • Batch 1, normal1.047×
  • Batch 1, sparse1.031×
  • Batch 1, wide1.021×

ResNet

  • Batch 1, normal0.311×
  • Batch 1, sparse0.319×
  • Batch 1, wide0.308×

MobileNet

  • Batch 1, normal0.529×
  • Batch 1, sparse0.533×
  • Batch 1, wide0.523×

Recurrent

  • Batch 1, normal1.002×
  • Batch 1, sparse1.001×
  • Batch 1, wide1.001×

Normalized MLP

  • Batch 1, normal0.892×
  • Batch 1, sparse0.959×
  • Batch 1, wide0.850×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction1.046×
  • compact-mixer on digits0.653×
  • compact-mixer on fashion1.020×
  • compact-mixer on mnist1.004×
  • compact-mlp on digits1.030×
  • compact-mlp on fashion1.270×
  • compact-mlp on mnist0.703×
  • compact-mobilenet on digits0.897×
  • compact-mobilenet on fashion1.650×
  • compact-mobilenet on mnist2.550×
  • compact-normalized_mlp on digits1.311×
  • compact-normalized_mlp on fashion0.835×
  • compact-normalized_mlp on mnist0.664×
  • compact-resnet on digits0.668×
  • compact-resnet on fashion1.165×
  • compact-resnet on mnist2.126×
  • appliance-gru on appliance-energy1.004×
  • mitbih-gru on mitbih0.999×
  • mnist-rows-lstm on mnist-rows1.001×
  • speech-commands-gru on speech-commands1.005×
  • speech-commands-lstm-gru on speech-commands1.001×
  • spoken-arabic-digit-lstm on spoken-arabic-digit0.998×
  • tiny-shakespeare-lstm on tiny-shakespeare0.998×
  • uci-har-gru on uci-har1.003×
  • uci-har-lstm on uci-har0.997×
  • uci-har-mixer on UCI-HAR0.964×
  • uci-har-mlp on UCI-HAR1.104×
  • uci-har-mobilenet on UCI-HAR0.454×
  • uci-har-normalized_mlp on UCI-HAR0.891×
  • uci-har-resnet on UCI-HAR0.282×
MLP2 runs

RunsGenerated 1.033× · real models 1.087×

MachineTime
  1. prod-amd64Seed 11563760321.008×Done1 attempt0.019× of the baseline kernels' time
  2. prod-arm64Seed 11563760321.114×Done1 attempt0.020× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal1.009×
  • Batch 1, sparse1.039×
  • Batch 1, wide1.106×

Real models

  • compact-mlp on digits1.426×
  • compact-mlp on fashion1.032×
  • compact-mlp on mnist0.938×
  • uci-har-mlp on UCI-HAR1.011×
ResNet2 runs

RunsGenerated 0.996× · real models 0.992×

MachineTime
  1. prod-amd64Seed 11563760321.000×Done1 attempt0.288× of the baseline kernels' time
  2. prod-arm64Seed 11563760320.988×Done1 attempt0.313× of the baseline kernels' time

By model and inputGuard at 1.100×

ResNet

  • Batch 1, normal0.997×
  • Batch 1, sparse0.997×
  • Batch 1, wide0.995×

Real models

  • compact-resnet on digits0.995×
  • compact-resnet on fashion0.994×
  • compact-resnet on mnist0.994×
  • uci-har-resnet on UCI-HAR0.985×
MobileNet2 runs

RunsGenerated 0.623× · real models 1.411×

MachineTime
  1. prod-amd64Seed 11563760320.941×Done1 attempt0.620× of the baseline kernels' time
  2. prod-arm64Seed 11563760320.934×Done1 attempt0.624× of the baseline kernels' time

By model and inputGuard at 1.100×

MobileNet

  • Batch 1, normal0.624×
  • Batch 1, sparse0.629×
  • Batch 1, wide0.616×

Real models

  • compact-mobilenet on digits1.091×
  • compact-mobilenet on fashion1.984×
  • compact-mobilenet on mnist3.256×
  • uci-har-mobilenet on UCI-HAR0.562×
Normalized MLP2 runs

RunsGenerated 1.012× · real models 1.011×

MachineTime
  1. prod-amd64Seed 11563760320.991×Done1 attempt0.010× of the baseline kernels' time
  2. prod-arm64Seed 11563760321.032×Done1 attempt0.013× of the baseline kernels' time

By model and inputGuard at 1.100×

Normalized MLP

  • Batch 1, normal0.974×
  • Batch 1, sparse1.091×
  • Batch 1, wide0.974×

Real models

  • compact-normalized_mlp on digits1.031×
  • compact-normalized_mlp on fashion0.988×
  • compact-normalized_mlp on mnist1.021×
  • uci-har-normalized_mlp on UCI-HAR1.006×
Mixer2 runs

RunsGenerated 1.068× · real models 1.027×

MachineTime
  1. prod-amd64Seed 11563760321.084×Done1 attempt0.005× of the baseline kernels' time
  2. prod-arm64Seed 11563760321.013×Done1 attempt0.005× of the baseline kernels' time

By model and inputGuard at 1.100×

Mixer

  • Batch 1, normal1.103×
  • Batch 1, sparse1.011×
  • Batch 1, wide1.100×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction1.236×
  • compact-mixer on digits0.904×
  • compact-mixer on fashion1.008×
  • compact-mixer on mnist1.004×
  • uci-har-mixer on UCI-HAR1.011×
Recurrent2 runs

RunsGenerated 15.250× · real models 17.224×

MachineTime
  1. prod-amd64Seed 115637603215.836×Done1 attempt7.870× of the baseline kernels' time
  2. prod-arm64Seed 115637603216.586×Done1 attempt8.243× of the baseline kernels' time

By model and inputGuard at 1.100×

Recurrent

  • Batch 1, normal15.717×
  • Batch 1, sparse14.691×
  • Batch 1, wide15.486×

Real models

  • appliance-gru on appliance-energy16.697×
  • mitbih-gru on mitbih18.904×
  • mnist-rows-lstm on mnist-rows18.036×
  • speech-commands-gru on speech-commands14.107×
  • speech-commands-lstm-gru on speech-commands16.059×
  • spoken-arabic-digit-lstm on spoken-arabic-digit19.643×
  • tiny-shakespeare-lstm on tiny-shakespeare15.178×
  • uci-har-gru on uci-har17.493×
  • uci-har-lstm on uci-har19.834×