Correct Wi-Fi spec and local-model performance in hardware docs - #273
Correct Wi-Fi spec and local-model performance in hardware docs#273yalexx wants to merge 1 commit into
Conversation
The hardware pages carried two inaccurate specs. Connectivity said "Wi-Fi 6". The Jetson Orin Nano Super dev kit we ship (945-13766-0005-000) bundles a Realtek RTL8822CE — 802.11 a/b/g/n/ac, dual-band 2x2, up to 867 Mbps, Bluetooth 5.0. Distributor listings for that exact part number specify an "802.11ac/abgn wireless network interface controller". So it is Wi-Fi 5, not Wi-Fi 6. Local model performance published ~3.5 tok/s for Llama 3.1 8B. Every row of that table works out to ~15-17 GB/s of effective memory bandwidth against the board's 102 GB/s, which is CPU-class — it was measured without CUDA offload. A shipped ClawBox sets MAXN + jetson_clocks at boot and runs Ollama with q8_0 KV cache and flash attention, i.e. the GPU path. Replaced with our own production measurement (2026-07-22 benchmark post): 1B ~45, 2B ~26, 3B ~23, 3.8B ~22 tok/s at ~11.4 W average, no throttling. 7-8B is listed honestly at ~10 tok/s and a tight fit in 8 GB. Same numbers now used on clawbox.com and in the support knowledge base. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
Warning Review limit reachedYou’ve reached a temporary PR review limit under our Fair Usage Limits Policy. Next review available in: 55 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (2)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
🦀 ClawReviewYour friendly reef crab, here with the lay of the land. Corrects two factual errors in the hardware docs: the bundled wireless adapter (RTL8822CE) is Wi-Fi 5/802.11ac, not Wi-Fi 6, and the local-model benchmark table is replaced with production-measured figures (Ollama, MAXN mode, 176 tegrastats samples) that better reflect what a shipped ClawBox actually achieves. Power draw figures are also updated to match the measured ~11 W typical / 19 W peak rather than the previous estimates. At a glance
Good to know
— ClawReview 🦀, your resident reef crab. Just orientation — CodeRabbit does the line-by-line, humans do the merge. Conventions: docs. |
1 participant
Two inaccurate specs on the hardware pages
Wi-Fi
The pages said Wi-Fi 6. The Jetson Orin Nano Super dev kit we ship (
945-13766-0005-000) bundles a Realtek RTL8822CE: 802.11 a/b/g/n/ac, dual-band 2×2, up to 867 Mbps, Bluetooth 5.0.Sources — distributor listings for that exact part number specify an "802.11ac/abgn wireless network interface controller" in the kit contents:
So it is Wi-Fi 5, not Wi-Fi 6.
Local model performance
The table published ~3.5 tok/s for Llama 3.1 8B. Multiply tok/s by Q4 weight size for every row and it lands at ~15–17 GB/s of effective memory bandwidth — against the board's 102 GB/s. That is CPU-class, so the table was measured without CUDA offload.
A shipped ClawBox does not run that way:
config/clawbox-performance.servicesets MAXN +jetson_clocksat bootscripts/optimize-ollama.shsetsOLLAMA_KV_CACHE_TYPE=q8_0,OLLAMA_FLASH_ATTENTION=1Replaced with our own production measurement (Ollama, warm runs,
eval_rate, tegrastats over 176 samples, GR3D 99%): 1B ~45 · 2B ~26 · 3B ~23 · 3.8B ~22 tok/s, ~11.4 W average, 61.8 °C peak, no throttling. 7–8B is now listed honestly at ~10 tok/s and a tight fit in 8 GB.For reference, NVIDIA's ~19 tok/s for Llama 3.1 8B on this board is MLC/INT4 — a runtime we do not ship.
Consistency
The same numbers are now on clawbox.com (#345, #346) and in the support knowledge base, so all three surfaces agree.
Still open
The 7–8B row is extrapolated from the measured bandwidth, not directly measured — worth one benchmark run on a live box to confirm, along with a 25 W vs MAXN comparison.
🤖 Generated with Claude Code