01
Embedded profiling
Profiles on Jetson Orin, IGX and DRIVE, where the same model behaves differently.
Edge · Edgyn
Edgyn specializes the Turbokern loop for Jetson, IGX and DRIVE. It optimizes for latency, power and thermal limits together and benchmarks on the real device, so the artifact meets the on-device budget.
Capabilities
01
Profiles on Jetson Orin, IGX and DRIVE, where the same model behaves differently.
02
Latency, power and temperature optimized together.
03
INT8 and INT4 tuned for edge tensor cores, with Quantra.
04
Every config measured on the physical module, not a cloud proxy.
05
DeepStream and Holoscan graph tuning for sensor workloads.
06
Deploy optimized models across device fleets.
The envelope
On a device, a faster model that overheats or drains the battery is not faster. Edgyn optimizes latency, power and temperature together.
Budget 30 ms per frame at 30 fps camera input.
Budget 15 W module power mode.
Budget below throttle point in a sealed enclosure.
How it works
Latency, power mode and thermal limit for the device.
Trace the model on the module itself.
Precision, kernels and engine flags under all three limits.
A TensorRT engine and a power-mode recipe.
Specification
| Item | Detail | Status |
|---|---|---|
| Devices | Jetson Orin Nano, NX, AGX | Pilot |
| Automotive | DRIVE platforms | Planned |
| Medical and industrial | IGX with Holoscan | Planned |
| Objectives | Latency, power, temperature | Pilot |
| Pipelines | DeepStream, Holoscan graph tuning | Research |
| Fleet deployment | Fleet Command rollout | Research |
Place in the loop
It reuses Profyle, Kernl and Quantra on embedded GPUs and feeds its device measurements back to Simyx.
FAQ
For the pilot we benchmark on our own Jetson modules. For custom carrier boards or enclosures, we run on yours.
Edgyn produces measurements and artifacts that can support your validation work. It does not certify anything, and certification scope for automotive and medical targets is still open.
Any nvpmodel power mode the module supports. You set the mode as part of the envelope.
Model optimization is in the pilot. DeepStream and Holoscan pipeline tuning is a research track.
Next step
Bring a perception model and its device budget to the pilot. Edgyn returns an engine and on-device measurements.