Edge AI Lab Brief: Dual-Thor Networking, Model Feasibility, and Test Automation

This week was less about producing another isolated benchmark and more about turning recent experiments into a repeatable workflow: decide whether a question is worth testing, measure the hardware boundary, and preserve the evidence in a form that can be rerun.

These are the three lessons worth carrying forward.

1. Port count is not system bandwidth

With four direct 10GbE links between two Jetson AGX Thor systems, every link individually reached roughly 9–10 Gbps. Under concurrent multi-link traffic, however, aggregate throughput stopped near 12 Gbps.

External interface count does not prove that the SoC contains the same number of independent internal data paths. A useful test sequence is: one link, multiple flows on one link, two concurrent links, and then all links together, while also recording CPU headroom, clocks, and power mode.

See the full measurements and deployment implications in Dual Jetson AGX Thor 4×10GbE Test.

2. “Can it run?” is really three questions

Before downloading a new model, answer at least three separate questions:

1. Do the weights and runtime state fit in available memory?
2. Does the inference stack support the architecture and quantization format?
3. If it loads, is performance useful for the intended workload?

The Jetson Thor and RTX 3090 comparison makes this separation important. Results collected at different precisions, runtimes, and memory architectures cannot be explained by one tokens-per-second number. A useful benchmark records quantization, context, concurrency, time to first token, and generation throughput together.

See Qwen3.8-27B on RTX 3090 vs Jetson Thor.

3. Automate discovery, not conclusions

The lab now has a minimal automation loop that discovers recent model candidates, prepares a review list, can trigger an RTX 3090 smoke test, and turns structured results into bilingual drafts.

Automation is effective for repetitive preparation. It should not promote “downloaded successfully” or “loaded once” into a performance conclusion. Thor testing, anomaly analysis, and final publication still require review, especially when framework support is incomplete.

Next validation queue

• Repeat the dual-Thor test across power modes and MTU settings.
• Add framework-support and estimated-memory gates to model discovery.
• Normalize benchmark fields across the RTX 3090 and Jetson Thor runners.
• Index failed experiments so the same dead ends are not repeated.

The point of a weekly brief is not to manufacture publishing cadence. It is to keep turning lab work into an evidence base that is searchable, comparable, and reproducible.