TB-3293 · REV C · Technical newsheet

Lab InstrumentationDevice profile

Remote Lab Streams a Jetson Nano to Any Browser at Under 100 ms

LiveBench streams a Jetson Nano 4GB to any browser with sub-100 ms terminal latency, adding remote flashing, power cycling and live power and thermal telemetry for distributed teams.

By Sophie Lindqvist4 min read796 words

Features

  • LiveBench reports terminal response latency under 100 ms on its remote lab setup
  • Demo hardware is an NVIDIA Jetson Nano 4GB developer kit accessed via browser
  • Benchmark covered YOLOv8n and YOLOv12n across input resolutions and execution formats
  • Remote relays and jumpers enable firmware flashing, power cycling and boot-recovery
  • Middleware keeps hardware off the public internet; each user works in an isolated sandbox
Remote Hardware Evaluation with LiveBench: Jetson Nano
Device photoRemote Hardware Evaluation with LiveBench: Jetson Nano — geralt / pixabay

LiveBench, a remote hardware evaluation lab from Tenxer Labs, delivers browser-based access to physical development boards with terminal response latency under 100 ms. The platform's demo lab exposes an NVIDIA Jetson Nano 4GB developer kit to remote users, complete with VS Code, firmware flashing, power cycling and live thermal and power telemetry.

The problem LiveBench targets is familiar to distributed embedded teams. Picture five developers in New York, Tokyo and Munich sharing one scarce evaluation kit (EVK) that sits in the New York office. The New York developers just power the board and work. The Tokyo developer who wants to test a model at the end of their workday has no way to power it up — and keeping it on 24/7, or asking a colleague to flip the switch, doesn't scale.

Even with the board powered, remote work isn't simple. Enterprise networks impose authentication and access controls, UART and CAN interfaces sit behind the OS, and someone still needs to recover the board after a failed flash.

Why benchmark on a legacy Jetson Nano?

Tenxer Labs deliberately chose the Jetson Nano 4GB, a legacy platform, because its hardware and software constraints make it a hard test case. If the remote experience works on older silicon, adapting it to newer hardware should be easier.

The lab setup pairs the board with a USB camera pointed at a display streaming test videos, so computer-vision applications run against a real-world video feed. A second camera gives remote users a live view of the physical hardware in operation.

What can a remote developer actually do?

LiveBench goes well beyond remote desktop. Depending on the EVK, lab designs integrate UART interfaces, USB devices and remotely controlled relays and jumpers, so users can:

  • Flash firmware or complete system images remotely
  • Access UART consoles to monitor boot logs and debug applications
  • Power the board on or off and trigger resets
  • Change boot and recovery configurations via remote relays and jumpers
  • Recover the board from failed flashes or unsuccessful boots

In short, developers interact at the bootloader and firmware level, as if the board sat on their desk. LiveBench's middleware provides controlled access without exposing hardware directly to the public internet. Each user works in an isolated sandbox with a pristine environment by default, so sharing a board doesn't mean sharing code.

Instrumentation beyond onboard sensors

The platform streams temperature, power consumption, CPU utilization and GPU utilization. On the Jetson Nano, several of these parameters come from existing system interfaces. Other lab setups add external sensors, oscilloscopes, digital multimeters and thermal cameras to capture measurements the board can't report itself, streamed alongside application performance data.

The experiment: YOLOv8 versus YOLOv12

Tenxer Labs benchmarked YOLOv8n and YOLOv12n (the "n" denotes the nano variant, the smallest model size) across different input resolutions and execution formats on the Jetson Nano. YOLOv8 is a popular edge-AI choice, while YOLOv12's attention-centric architecture carries different computational requirements. Developers wrote and executed the application in VS Code on the remote board, iterating as needed.

For input, the USB camera watched a display streaming preloaded videos, and users could also stream their own laptop webcam to that display. Test material spanned objects from tiny to large.

What gets measured, and why

The team monitored frame rate (FPS), inference time, and total processing time — the full pipeline including preprocessing, inference and postprocessing. That separation matters: operations like Non-Maximum Suppression (NMS), the step that filters overlapping detections, can add substantial CPU overhead even when model inference is fast.

Alongside those, they tracked CPU and GPU utilization, RAM usage, the thermal profile under sustained workloads, and power consumption. The results, detailed in Tenxer Labs' previous blog posts, underscore that the right combination of EVK, peripherals, model and model format depends on whether an application prioritizes accuracy — favoring higher resolutions — or speed and power, favoring lower resolutions or optimized formats like TensorRT.

Does it feel local?

Latency is what makes or breaks remote development, and Tenxer Labs reports terminal response latency under 100 ms in its setup, though actual latency varies with the user's last-mile connectivity. Developers who tried the setup reportedly didn't believe the Jetson Nano was in a remote lab. The platform also supports benchmarks across PyTorch, TensorRT and ONNX inference frameworks.

For the Tokyo developer in the opening scenario, the dependency on New York disappears: they can power the board, write and run applications, flash firmware and collect power and thermal data independently.

The open question for engineering managers: does controlled, sandboxed remote access finally remove scarce-hardware bottlenecks from your team's schedule, and does a middleware-mediated lab satisfy your security policies?

The demo lab is accessible at livebench.tenxerlabs.com/app/Jetson-001.

via hackster.io (Source)

Filed under

  • remote-lab
  • jetson-nano
  • edge-ai
  • hardware-benchmarking
  • yolo
Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering regulation, policy and their impact on Instruments, test and measurement at Testbench Report.

110 articles

Application notes

« Previous articleNext article »