Conjectures.io
Tensor kernels

Entry #66

Completed
Hotkey
5DvXV7…peR7qk
Sent
11 Nov 2026, 19:14 UTC
Source
9f624c1c7110

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 runs63.147×Generated 67.220×, real models 59.320×Above the barGain lower bound -6750.1%Guard passed
MLPvs #44Done2 of 2 runs0.977×Generated 0.962×, real models 0.992×Above the barGain lower bound -22.3%Guard passed
ResNetvs #51Done2 of 2 runs114.850×Generated 137.999×, real models 95.584×Above the barGain lower bound -12921.5%Guard passed
MobileNetvs #53Done2 of 2 runs107.830×Generated 141.437×, real models 82.209×Above the barGain lower bound -14260.9%Guard passed
Normalized MLPvs #48Done2 of 2 runs49.130×Generated 64.587×, real models 37.372×Above the barGain lower bound -6007.1%Guard passed
Mixervs #47Done2 of 2 runs196.401×Generated 205.882×, real models 187.358×Above the barGain lower bound -21532.5%Guard passed
Recurrentvs #52Done2 of 2 runs777.708×Generated 680.579×, real models 888.698×Above the barGain lower bound -80432.0%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 67.220× · real models 59.320×

MachineTime
  1. prod-amd64Seed 383746147065.133×Done1 attempt0.654× of the baseline kernels' time
  2. prod-arm64Seed 383746147061.221×Done1 attempt0.682× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal0.984×
  • Batch 1, sparse1.088×
  • Batch 1, wide1.219×

Mixer

  • Batch 1, normal187.775×
  • Batch 1, sparse186.642×
  • Batch 1, wide213.666×

ResNet

  • Batch 1, normal114.648×
  • Batch 1, sparse135.168×
  • Batch 1, wide115.118×

MobileNet

  • Batch 1, normal92.219×
  • Batch 1, sparse93.647×
  • Batch 1, wide93.421×

Recurrent

  • Batch 1, normal707.848×
  • Batch 1, sparse659.629×
  • Batch 1, wide656.080×

Normalized MLP

  • Batch 1, normal63.719×
  • Batch 1, sparse65.683×
  • Batch 1, wide65.817×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction229.546×
  • compact-mixer on digits210.647×
  • compact-mixer on fashion190.576×
  • compact-mixer on mnist180.331×
  • compact-mlp on digits1.099×
  • compact-mlp on fashion1.126×
  • compact-mlp on mnist1.584×
  • compact-mobilenet on digits64.894×
  • compact-mobilenet on fashion82.260×
  • compact-mobilenet on mnist77.835×
  • compact-normalized_mlp on digits73.968×
  • compact-normalized_mlp on fashion22.509×
  • compact-normalized_mlp on mnist23.424×
  • compact-resnet on digits71.236×
  • compact-resnet on fashion120.241×
  • compact-resnet on mnist120.771×
  • appliance-gru on appliance-energy763.115×
  • mitbih-gru on mitbih891.405×
  • mnist-rows-lstm on mnist-rows730.611×
  • speech-commands-gru on speech-commands688.009×
  • speech-commands-lstm-gru on speech-commands867.489×
  • spoken-arabic-digit-lstm on spoken-arabic-digit1130.465×
  • tiny-shakespeare-lstm on tiny-shakespeare829.741×
  • uci-har-gru on uci-har834.946×
  • uci-har-lstm on uci-har1094.018×
  • uci-har-mixer on UCI-HAR202.336×
  • uci-har-mlp on UCI-HAR1.013×
  • uci-har-mobilenet on UCI-HAR74.299×
  • uci-har-normalized_mlp on UCI-HAR20.137×
  • uci-har-resnet on UCI-HAR79.884×
MLP2 runs

RunsGenerated 0.962× · real models 0.992×

MachineTime
  1. prod-amd64Seed 38374614700.943×Done1 attempt0.015× of the baseline kernels' time
  2. prod-arm64Seed 38374614701.012×Done3 attempts0.018× of the baseline kernels' time

By model and inputGuard at 1.100×

MLP

  • Batch 1, normal0.965×
  • Batch 1, sparse0.959×
  • Batch 1, wide0.918×

Real models

  • compact-mlp on digits0.566×
  • compact-mlp on fashion0.969×
  • compact-mlp on mnist1.876×
  • uci-har-mlp on UCI-HAR0.944×
ResNet2 runs

RunsGenerated 137.999× · real models 95.584×

MachineTime
  1. prod-amd64Seed 3837461470108.561×Done1 attempt0.996× of the baseline kernels' time
  2. prod-arm64Seed 3837461470121.502×Done1 attempt1.020× of the baseline kernels' time

By model and inputGuard at 1.100×

ResNet

  • Batch 1, normal162.431×
  • Batch 1, sparse134.497×
  • Batch 1, wide116.112×

Real models

  • compact-resnet on digits71.972×
  • compact-resnet on fashion119.815×
  • compact-resnet on mnist120.396×
  • uci-har-resnet on UCI-HAR80.399×
MobileNet2 runs

RunsGenerated 141.437× · real models 82.209×

MachineTime
  1. prod-amd64Seed 383746147088.872×Done1 attempt1.181× of the baseline kernels' time
  2. prod-arm64Seed 3837461470130.832×Done1 attempt1.170× of the baseline kernels' time

By model and inputGuard at 1.100×

MobileNet

  • Batch 1, normal133.675×
  • Batch 1, sparse126.384×
  • Batch 1, wide130.709×

Real models

  • compact-mobilenet on digits65.421×
  • compact-mobilenet on fashion95.305×
  • compact-mobilenet on mnist93.500×
  • uci-har-mobilenet on UCI-HAR78.348×
Normalized MLP2 runs

RunsGenerated 64.587× · real models 37.372×

MachineTime
  1. prod-amd64Seed 383746147055.902×Done1 attempt0.572× of the baseline kernels' time
  2. prod-arm64Seed 383746147043.178×Done1 attempt0.568× of the baseline kernels' time

By model and inputGuard at 1.100×

Normalized MLP

  • Batch 1, normal65.685×
  • Batch 1, sparse81.661×
  • Batch 1, wide74.540×

Real models

  • compact-normalized_mlp on digits69.295×
  • compact-normalized_mlp on fashion32.743×
  • compact-normalized_mlp on mnist38.762×
  • uci-har-normalized_mlp on UCI-HAR22.179×
Mixer2 runs

RunsGenerated 205.882× · real models 187.358×

MachineTime
  1. prod-amd64Seed 3837461470204.875×Done1 attempt0.924× of the baseline kernels' time
  2. prod-arm64Seed 3837461470188.278×Done1 attempt0.917× of the baseline kernels' time

By model and inputGuard at 1.100×

Mixer

  • Batch 1, normal200.360×
  • Batch 1, sparse198.156×
  • Batch 1, wide230.386×

Real models

  • appliance-energy-mixer on Appliance Energy Prediction267.227×
  • compact-mixer on digits144.568×
  • compact-mixer on fashion180.361×
  • compact-mixer on mnist181.522×
  • uci-har-mixer on UCI-HAR182.528×
Recurrent2 runs

RunsGenerated 680.579× · real models 888.698×

MachineTime
  1. prod-amd64Seed 3837461470813.508×Done1 attempt8.013× of the baseline kernels' time
  2. prod-arm64Seed 3837461470743.483×Done1 attempt8.077× of the baseline kernels' time

By model and inputGuard at 1.100×

Recurrent

  • Batch 1, normal722.652×
  • Batch 1, sparse679.977×
  • Batch 1, wide670.391×

Real models

  • appliance-gru on appliance-energy804.178×
  • mitbih-gru on mitbih904.585×
  • mnist-rows-lstm on mnist-rows747.820×
  • speech-commands-gru on speech-commands708.041×
  • speech-commands-lstm-gru on speech-commands879.765×
  • spoken-arabic-digit-lstm on spoken-arabic-digit1165.225×
  • tiny-shakespeare-lstm on tiny-shakespeare846.827×
  • uci-har-gru on uci-har860.976×
  • uci-har-lstm on uci-har1201.079×