Hosted evaluation boards
Real E-series boards with exact revisions, SDK and firmware versions. Availability and compatibility are shown before you build.
E1
E2
E4Veyra AI Hub
Test your model on real Veyra silicon.
Upload an ONNX or TFLite model, pick a hosted E-series board, and get measured latency and memory — before you order hardware.
3 hosted boardsReady-to-run examplesExportable results
From model to measured result
One guided path for every project. Each step checks what the next one needs.
- 1Model
Upload ONNX / TFLite or pick a Model Zoo example.
- 2Target
Choose an exact hosted board. Compatibility is checked.
- 3Build
Compile with a supported FP16 or INT8 profile.
- 4Inputs
Use bundled samples or your own test data.
- 5Run
Execute on the board and follow the live log.
- ✓Results
Latency, memory and output — then compare and export.
What you get
E1E2E4Model Zoo with ready-to-run examples
Start without uploading anything: validated vision, audio and time-series models with bundled sample inputs.
VisionAudioNumeric
Guided build profiles
Only the precisions a board supports. INT8 quantization comes with a matching calibration set.
Balanced FP16Fast INT8
Results with full context
Median and p95 warm latency, peak runtime memory and the inference output — tied to the exact board, build and inputs.
Median12.6 ms
p9514.1 ms
Peak memory38 MiB
Compare runs and take them home
Put two runs side by side to see what changed, then export a package to reproduce the run on your own bench.
12.6 ms8.4 ms−33.3%
Veyra Assistant at every step
Explains errors and compatibility, cites the exact Veyra document, and asks before changing anything.
E1 allows 16 MiB runtime; this model needs 26 MiB at INT8. Try VY-E2 instead?E1 Evaluation Profile v2.1 §4
Ready to evaluate?
Sign in to Veyra AI Hub with the demo account and open the Warehouse Vision project.
Latency numbers on this page are illustrative marketing figures.