Predict · Simyx

Predict the best optimization before you burn the compute

Simyx is a learned performance twin of NVIDIA GPUs. It estimates the latency, throughput and accuracy of candidate optimizations before they run, so every other product benchmarks only the most promising few.

Prediction batchExample
Proposed by
Kernl
Candidates
212
Sent to GPU
18
GPU-hours saved
~50 of 64
Model
H100, attention class
Example. Values are illustrative.

Capabilities

What Simyx does

01

Performance predictors

Latency and throughput models per workload class and GPU architecture.

02

Accuracy predictor

Estimates quality loss of precision changes before they run.

03

Search-budget control

Caps GPU spend per run and allocates it to the best-ranked candidates.

04

Active learning

Chooses some candidates to run for what the model will learn, not only for speed.

05

What-if estimates

Estimate a change or a new GPU before you buy or build anything.

Planned

06

Capacity planning

Fleet-level simulation for GPU planning.

Planned

Guided search

Fewer benchmarks for the same win

Simyx estimates each candidate's speed and accuracy before it runs. Only the top-ranked candidates use GPU time.

GPU-hours to reach the same kept resultIllustrative
Brute-forceWith SimyxIllustrative
Kernel search
64 h
14 h
Quantization search
40 h
12 h
Serving search
28 h
10 h
Illustrative planning figures, not measured.
Predictor checkDesign targets
Candidates scored
212
Benchmarked
Top 18
Rank agreement
Target ≥ 0.8
Missed winners
Logged and re-learned
Explore share
10% random
Design targets for the v1 predictor.

How it works

Inside a Simyx run

  1. 01Milliseconds

    Featurize

    Turn each candidate and workload into trace-level features.

  2. 02Milliseconds

    Predict

    Score speed and accuracy with uncertainty.

  3. 03Instant

    Select

    Send the top-k and a few exploratory picks to the GPU.

  4. 04Continuous

    Learn

    Fold measured results back into the model.

Specification

Inputs, outputs and hardware

ItemDetailStatus
Used byProfyle, Kernl, Quantra, Servex, EdgynPilot
PredictsLatency, throughput; accuracy nextPilot · Next
Training dataTurbokern optimization-trajectory datasetPilot
ComputeRAPIDS on GPU for training and analyticsPilot
Standalone what-ifEstimates without a full runPlanned
Capacity planningFleet right-sizing simulationResearch

FAQ

Questions engineers ask

Is a prediction ever used instead of a measurement?

No. Simyx only decides what to measure. Nothing is kept without a real benchmark on the target GPU.

What happens on a brand-new GPU?

Predictions start less certain. Simyx sends more candidates to the GPU until it has learned the new architecture, which costs more compute early on.

Does my data train a shared model?

Only anonymous performance features, and only if you allow pooling. Weights, prompts and eval data are never used.

Can I use Simyx on its own?

A standalone what-if and capacity-planning tier is planned after the pilot.

Next step

Spend GPU-hours where they pay

In the pilot every run reports how many candidates Simyx pruned and what that saved.