From a9a49acee99edb69971ff00174b15a80432f88b4 Mon Sep 17 00:00:00 2001 From: Yanko Atanasov Aleksandrov Date: Mon, 27 Jul 2026 16:26:24 +0300 Subject: [PATCH] docs: correct Wi-Fi spec and local-model performance to measured values MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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 --- docs-site/hardware/clawbox-connect.mdx | 19 +++++++++++++------ docs-site/hardware/requirements.mdx | 21 +++++++++++++-------- 2 files changed, 26 insertions(+), 14 deletions(-) diff --git a/docs-site/hardware/clawbox-connect.mdx b/docs-site/hardware/clawbox-connect.mdx index 7ce66d87..499bfce2 100644 --- a/docs-site/hardware/clawbox-connect.mdx +++ b/docs-site/hardware/clawbox-connect.mdx @@ -19,9 +19,9 @@ The entry-tier ClawBox: a compact, low-power, always-on AI assistant for your de | GPU | 1024-core NVIDIA Ampere, 625 MHz | | Memory | 8 GB LPDDR5, 102 GB/s bandwidth (unified CPU/GPU) | | Storage | 512 GB NVMe SSD (PCIe Gen3 x4, ~2,100 MB/s) | -| Connectivity | Wi-Fi 6 + Bluetooth 5.0 + Gigabit Ethernet | +| Connectivity | Wi-Fi 5 (802.11ac, dual-band 2×2, up to 867 Mbps) + Bluetooth 5.0 + Gigabit Ethernet | | Ports | USB-A, USB-C, HDMI, Ethernet, microSD | -| Power draw | 7 W idle · 15 W typical · 25 W max (20 W USB-C PSU included) | +| Power draw | 7 W idle · ~11 W typical / 19 W peak measured · 25 W max (20 W USB-C PSU included) | | Size | 100 × 79 × 31 mm, carbon-color case (~260 g) | | OS | Ubuntu 22.04 LTS + OpenClaw pre-installed | @@ -35,10 +35,17 @@ The entry-tier ClawBox: a compact, low-power, always-on AI assistant for your de | Model | Speed | Quality | |---|---|---| -| Llama 3.1 8B (Q4) | ~3.5 tok/s | Excellent | -| Mistral 7B (Q4) | ~4.2 tok/s | Excellent | -| Phi-3 Mini (Q4) | ~6.8 tok/s | Good | -| Gemma 2B (Q4) | ~9.1 tok/s | Fast | +| Llama 3.2 1B (Q4) | ~45 tok/s | Fast | +| Gemma 2 2B (Q4) | ~26 tok/s | Good | +| Qwen2.5 3B (Q4) | ~23 tok/s | Excellent | +| Phi-3 Mini 3.8B (Q4) | ~22 tok/s | Excellent | +| 7–8B class (Q4) | ~10 tok/s | Tight fit | + +Measured on a production ClawBox with Ollama (warm runs, `num_predict=200`, Ollama's own +`eval_rate`; power and thermals from `tegrastats` over 176 samples): ~11.4 W average, +19.2 W peak, 61.8 °C peak, no throttling. Most people read at 5–8 tok/s, so the 1–4B +class generates faster than you can read. Full methodology: +[We Benchmarked a Production ClawBox](https://clawbox.com/blog/2026-07-22-jetson-orin-nano-llm-benchmark-real-clawbox-numbers). See full requirements and tiers in [Hardware Requirements](/hardware/requirements). diff --git a/docs-site/hardware/requirements.mdx b/docs-site/hardware/requirements.mdx index 3ca2881c..5e940fd3 100644 --- a/docs-site/hardware/requirements.mdx +++ b/docs-site/hardware/requirements.mdx @@ -48,7 +48,7 @@ This page covers what you actually need — from bare minimum to running local A ### Power User — local AI models - **Specs:** 4+ cores, 16 GB+ RAM, 100 GB+ SSD, 25 Mbps+, NVIDIA GPU, Linux (Ubuntu/Debian) -- **Can do:** everything above, plus local LLMs (7B–13B), local image generation, on-device speech recognition, fully offline operation, multiple browser instances +- **Can do:** everything above, plus local LLMs (1B–8B; 1–4B is the sweet spot on ClawBox Connect), local image generation, on-device speech recognition, fully offline operation, multiple browser instances - **Can't do:** 70B+ models need more VRAM → see [ClawBox Workstation](/hardware/clawbox-workstation) - **Examples:** **ClawBox Connect** (67 TOPS) · gaming PC with GPU · **ClawBox Workstation** (DGX Spark) @@ -68,12 +68,17 @@ Running OpenClaw on dedicated hardware vs a shared VPS or your daily-use laptop: | Model | Speed | Quality | |---|---|---| -| Llama 3.1 8B (Q4) | ~3.5 tok/s | Excellent | -| Mistral 7B (Q4) | ~4.2 tok/s | Excellent | -| Qwen2.5 7B (Q4) | ~4.0 tok/s | Excellent | -| Phi-3 Mini (Q4) | ~6.8 tok/s | Good | -| Gemma 2B (Q4) | ~9.1 tok/s | Fast | -| TinyLlama 1B (Q4) | ~15.3 tok/s | Basic | +| Llama 3.2 1B (Q4) | ~45 tok/s | Fast | +| Gemma 2 2B (Q4) | ~26 tok/s | Good | +| Qwen2.5 3B (Q4) | ~23 tok/s | Excellent | +| Phi-3 Mini 3.8B (Q4) | ~22 tok/s | Excellent | +| 7–8B class (Q4) | ~10 tok/s | Tight fit | + +Measured on a production ClawBox with Ollama (warm runs, `num_predict=200`, Ollama's own +`eval_rate`; power and thermals from `tegrastats` over 176 samples): ~11.4 W average, +19.2 W peak, 61.8 °C peak, no throttling. Most people read at 5–8 tok/s, so the 1–4B +class generates faster than you can read. Full methodology: +[We Benchmarked a Production ClawBox](https://clawbox.com/blog/2026-07-22-jetson-orin-nano-llm-benchmark-real-clawbox-numbers). For larger models, use your own cloud API key (Claude, GPT, Gemini) — see @@ -84,6 +89,6 @@ see [ClawBox Workstation](/hardware/clawbox-workstation) (128 GB unified memory, ## Skip the DIY - ClawBox ships with Recommended+ specs, OpenClaw pre-installed, dual-band Wi-Fi and Bluetooth. + ClawBox ships with Recommended+ specs, OpenClaw pre-installed, dual-band Wi-Fi 5 (802.11ac) and Bluetooth 5.0. Manual DIY setup is typically 2–4+ hours (plus waiting for hardware).