Board: NVIDIA Jetson Orin Nano Super (8GB, JetPack 6.2, R36.4.3)
CUDA: 12.6, Compute Capability 8.7
Storage: 57GB eMMC (root) + 938GB NVMe
Runtime: llama.cpp (CUDA build, fbd441c37)
Date: 2026-04-02
Configuration: Headless (no USB devices, no video, no cameras)
Model Test 15W 25W MAXN 15W→MAXN Gain TinyLlama 1.1B (Q4_K_M, 636MB)Prompt (pp512) 1,026 1,398 1,506 +47% Generation (tg128) 35.3 47.8 51.1 +45% Phi-3 3.8B (Q4, 2.23GB)Prompt (pp512) 350 492 537 +53% Generation (tg128) 12.2 16.0 16.9 +39% Llama 3.2 3B (Q4_K_M, 1.87GB)Prompt (pp512) 479 650 713 +49% Generation (tg128) 13.3 17.2 18.4 +38%
Test 15W Mode 25W Mode MAXN Idle 5.2W 5.5W 9.1W CPU-only stress (6 cores)7.6W 8.0W 9.1W TinyLlama inference 10.3W avg / 11.1W peak 12.7W avg / 13.6W peak 14.0W avg / 15.0W peak Phi-3 3.8B inference 11.7W avg / 12.5W peak 14.6W avg / 15.3W peak 16.0W avg / 16.9W peak Llama 3.2 3B inference 11.2W avg / 11.9W peak 13.4W avg / 14.4W peak 15.0W avg / 16.0W peak Combined CPU+GPU 6.8W avg / 14.5W peak 8.4W avg / 17.0W peak 9.9W avg / 19.3W peak
Subsystem Avg Power Peak Power Notes Idle baseline 5.7W 5.7W MAXN, no load, post-soak NVMe I/O only (fio seq+rand)7.0W 7.6W +1.9W over idle Network only (iperf3 GbE flood)6.0W 6.5W +0.3-0.8W over idle GPU inference only (Phi-3)16.0W 16.9W Dominates power budget
Mode Avg Power Peak Power Max Tj 15W 8.3W 14.9W 60.5C 25W 8.2W 17.4W 62.8C MAXN 8.5W 20.0W 64.5C
Metric Value Peak power 20.4W Sustained power (under full load) 17.2-20.2W Tj at start 51.2C Tj peak (at ~4 min) 71.0C Tj stabilized 67-68C Tj after cooldown 49.7C GPU clock start 1015 MHz GPU clock end 1013 MHz Thermal throttling None Performance degradation None
Run Phi-3 pp1024 (t/s) Phi-3 tg256 (t/s) 1 473.92 16.67 2 473.31 16.69 3 474.01 16.51 4 474.48 16.95 5 476.90 16.96 6 475.50 16.98 7 477.03 16.98 8 476.61 16.97
Performance actually improved slightly as components warmed up. Zero degradation.
Mode Bogo ops/s CPU Freq 15W 1,052 1497 MHz 25W 944 1344 MHz MAXN 1,215 1728 MHz
Note: 25W mode scored lower than 15W on CPU — 25W sets CPU to 1344 MHz vs 15W’s 1497 MHz. The 25W budget goes more to GPU clocks (914 MHz vs 611 MHz).
Test Throughput Sequential write (1M blocks) 915 MB/s Sequential read (1M blocks) 1,456 MB/s Random 4K read (4 jobs) 144 MB/s Random 4K write (4 jobs) 145 MB/s
Direction Throughput Upload (4 streams) 457 Mbits/sec Download (4 streams) 522 Mbits/sec
Mode GPU Clock 15W 611 MHz 25W 911-917 MHz MAXN 1011-1019 MHz
Observation Detail Max junction temp 71.0C (MAXN 5-min soak, all subsystems) Thermal throttle threshold ~97C (Orin Nano) Headroom ~26C below throttle point GPU clock stability Rock solid across all tests — no frequency drops Thermal stabilization Peaks at ~70C after 4 min, settles to 67-68C Cooldown Returns to 50C within ~5 min after load stops
Test Start Tj End Tj Max Tj GPU Clock (start/end) 15W TinyLlama 50.1C 52.7C 53.3C 611 / 611 15W Phi-3 51.9C 56.3C 57.9C 611 / 611 15W Llama 3.2 56.0C 57.8C 58.8C 611 / 611 25W TinyLlama 50.3C 54.4C 54.4C 916 / 911 25W Phi-3 51.9C 58.5C 59.3C 914 / 914 25W Llama 3.2 56.1C 58.8C 60.4C 912 / 914 MAXN TinyLlama 51.5C 56.2C 56.2C 1018 / 1011 MAXN Phi-3 50.7C 60.3C 60.3C 1012 / 1017 MAXN Llama 3.2 56.1C 60.1C 61.9C 1015 / 1017
Scenario Power Draw Idle (MAXN) 5.7W Typical inference load 15-17W Absolute worst case (all subsystems, MAXN) 20.4W
Cluster Size Min PSU (at measured max) Recommended (20% margin) 1 node 21W 25W 2 nodes 41W 50W 4 nodes 82W 100W 5 nodes 102W 125W
True headless max power: 20.4W at MAXN — even with CPU, GPU, NVMe, and network all saturated simultaneouslyMAXN gives ~40-53% more inference throughput over 15W for only ~5W more actual drawThe board never reaches its 25W MAXN envelope — the 8GB Orin Nano simply can’t draw that much without USB/video/camera peripheralsZero thermal throttling — peaked at 71C under sustained full load, 26C below the throttle pointZero performance degradation — inference speed was rock steady across 8 consecutive runs during the 5-min soakGeneration speed is memory-bandwidth limited — 12-18 tok/s on 3B models regardless of how hard you push the GPUNVMe adds ~2W, network adds <1W to the power budget — the GPU dominatesBudget 25W per node with 20% margin for PSU sizing in a headless cluster
Test Avg Power Peak Power Peak Current Status Idle (MAXN) 6.75W 6.75W 0.56A ✅ Safe CPU stress (6 cores)~10.0W 10.2W 0.85A ✅ Safe TinyLlama 1.1B 14.9W 17.5W 1.46A ⚠️ Borderline Llama 3.2 3B 18.0W 19.0W 1.58A ❌ Over 1.5A spec Phi-3 3.8B 19.3W 20.2W 1.68A ❌ Over 1.5A spec Kitchen sink (all subsystems)11.3W 21.7W 1.81A ❌ 21% over spec
Metric Value Min power 8.3W Avg power 20.0W Peak power 21.7W Peak current 1.81A at 12V Tj start 53.4C Tj avg 70.8C Tj max 73.0C Tj end 71.9C Throttling None Brownout / crash None PDU port temp Cool to touch
Test Avg Power Peak Power Peak Current Status Idle (MAXN) 6.76W 6.76W 0.57A ✅ Safe CPU stress (6 cores)~10.3W 10.25W 0.85A ✅ Safe TinyLlama 1.1B — 17.5W 1.46A ⚠️ Borderline Llama 3.2 3B — 18.9W 1.57A ❌ Over 1.5A spec Phi-3 3.8B — 20.1W 1.68A ❌ Over 1.5A spec
Metric Value Peak power 21.73W Peak current 1.81A at 12V Max Tj 73.3C Throttling None Brownout / crash None PDU port temp Cool to touch
Metric Port 1 Port 2 Port 3 Peak power 21.7W 21.7W 21.73W Peak current 1.81A 1.81A 1.81A Max Tj 73.0C 73.2C 73.3C Throttling None None None Brownout / crash None None None Port temp Cool Cool Cool
Ports are rated 1.5A but held 1.81A sustained for 5 minutes without failure across all 3 ports
All PDU components remained cool — suggests conservative rating or low-resistance implementation
Inference performance was rock solid throughout the soak on all ports
Results are highly consistent port-to-port (<1% variance on peak power/current, <0.3C on Tj)
For dev/test clusters: Existing PDU appears viable — ports are clearly overbuilt vs. their 1.5A specFor production clusters: Further validation recommended, direct wiring to PSU is safestNVMe auto-mounts on boot via /etc/fstab with nofail ✅
Configuration: PicoCluster Claw acrylic case, 80x25mm case fan, PDU, 50W PSU
Nodes: clusterclaw (RPi5 8GB, Raspbian) + clustercrush (Orin Nano Super 8GB, MAXN)
Measurement: Kill-A-Watt at wall, tegrastats on Orin, vcgencmd PMIC on RPi5
Test Out-of-Case Tj Peak In-Case Tj Peak Improvement CPU-only stress (5 min) 64.5C 54.4C -10.1C Kitchen sink + GPU inference (5 min) 71.0C 65.4C -5.6C
Case fan provides significant cooling benefit. No throttling in either configuration.
State Temperature Idle 31.2C CPU + matrix + I/O stress (peak) 67.5C Post-cooldown 33.4C
State Kill-A-Watt Notes Both idle 13-14W Baseline Both under CPU/IO stress (no GPU inference)21-22W RPi5 headroom-limited Both max + GPU inference ~35W (estimated)Orin peaks at 21.6W SoC
Component Idle Typical (agent workload) Peak (kitchen sink) RPi5 (clusterclaw) ~3W ~4W 7.5W Orin Nano (clustercrush) ~6W ~10W 20W 12V fan ~2W ~2W ~2W PDU ~1W ~1W ~1W PSU efficiency loss (~15%) ~2W ~2.5W ~4.5W Total per pair ~14W ~19.5W ~35W
Metric Value VDD_IN peak 21.6W VDD_IN sustained 20.8-21.5W VDD_CPU_GPU_CV peak 11.7W VDD_SOC peak 4.5W GPU utilization 92-99% @ 1014-1017 MHz CPU utilization 94-100% @ 1728 MHz (6 cores) Tj peak (in case) 65.4C Tj at start 49.9C Throttling None
Real-world OpenClaw agent loop is bursty (inference → wait for browser → observe → repeat):
Phase Duration Orin Power Inference (token generation) ~12s ~16W Idle (waiting for browser action) ~25s ~6W Weighted cycle average ~37s ~9-10W
Agent is idle (waiting for user commands) 90%+ of the time.
US average electricity: $0.16/kWh. PSU: 50W (15W headroom at peak).
Scenario Watts kWh/month kWh/year $/month $/year Idle (waiting for tasks)14W 10.1 122.6 $1.61 $19.62 Active agent workload 20W 14.4 175.2 $2.30 $28.03 Peak (all subsystems maxed)35W 25.2 306.6 $4.03 $49.06 Realistic blend (90% idle, 10% active)15W 10.8 131.4 $1.73 $21.02
Device Typical Power Annual Cost PicoCluster Claw cluster 15W ~$21/year Desktop PC (always on) ~80W ~$112/year Single GPU server (RTX 4090) ~300W+ ~$420+/year 60W light bulb (always on) 60W ~$84/year
Your own private AI agent for under $2/month.
Orin out-of-case benchmarks: /mnt/nvme/results/ on clustercrush
Orin in-case benchmarks: /home/picocluster/results-incase/ on clustercrush
RPi5 benchmarks: /home/picocluster/results/ on clusterclaw
Date: 2026-05-19Hardware: Jetson Orin Nano Super 8GB (MAXN, Ollama CUDA)Gateway: OpenClaw 2026.5.12 via LiteLLM proxyContext: ~9,250 chars per request (optimized lean profile for local models)
The OpenClaw benchmark measures real agent capability across four tiers of increasing complexity — raw inference through multi-step tool chaining — using the live gateway, not a raw API.
Tier Tests What it measures T1 ping, math, JSON Raw inference, instruction following T2 list files, read, write Single tool call with verification T3 write→read chain, compute→write Multi-step tool chaining T4 write→edit→verify, read→summarize→write Harder reasoning + tool use
Model T1 T2 T3 T4 Score Avg latency granite4.1:8b 3/3 3/3 2/2 2/2 10/10 16,761 ms nemotron-3-nano:4b 3/3 3/3 2/2 1/2 9/10 38,895 ms qwen3.5:4b 2/3 3/3 2/2 1/2 8/10 45,824 ms llama3.1:8b 3/3 2/3 1/2 0/2 6/10 27,144 ms llama3.2:3b 3/3 1/3 0/2 0/2 5/10 8,957 ms
granite4.1:8b is the validated default: perfect score across all tiers.
nemotron-3-nano:4b at 9/10 is the recommended fast alternative — 4B parameters with near-8B tool reliability.
Validated threshold: T2+ required for agent work. granite, nemotron, and qwen3.5 all qualify.
Script: bench-claw.py (30 tasks across cap and workflow suites)Date: May 2026SMEEP: Small Model Experimentation and Evaluation Project
30 tasks that test real-world agent behavior: natural-language exec, web research, scheduling, file operations, and multi-tool chains. Every task is something a real user would type.
Suite Tasks What it tests cap14 Structured capability: exec, web, scheduling, chains workflow16 Natural language: commands a real user would type
Model Cap (14) Workflow (16) Total Avg task time qwen2.5:7b — — 14/30 37s granite4.1:8b — — 12/30 68s nemotron-3-nano:4b — — 12/30 57s llama3.1:8b — — 11/30 30s qwen3.5:4b — — 9/30 63s
Cap/workflow breakdown pending full bench analysis. Run in progress for phi4-mini:3.8b.
Tool context injection triples granite’s score. Without a structured context block
documenting available tools, granite4.1:8b scored 4/30. With it: 12/30. Granite needs
explicit tool documentation; it won’t infer from context what’s available.
7B is a meaningful step. qwen2.5:7b at 14/30 leads all models — including 8B models.
Natural-language command interpretation is where the size difference shows up: “How’s the
memory looking?” maps correctly to cat /proc/meminfo without a hint.
Mac 14B comparison (early data): mac/qwen2.5:14b scored 20/30 on the same 30-task suite.
14B closes the gap on multi-tool chains and complex scheduling. Full results pending.
git clone https://github.com/picocluster/PicoCluster-Claw
# Run on any Ollama host with OpenClaw running
python3 scripts/bench-claw.py \
--models granite4.1:8b,qwen2.5:7b,nemotron-3-nano:4b \
--output results/my-bench.json
Full methodology: docs/research/benchmarks/methodology.md
Earlier runs with the default OpenClaw context showed granite scoring 5/10 and qwen scoring 3/10 — small models were overwhelmed before they could reason about the actual task.
Key changes:
Change Saved Impact Replaced full workspace file templates with minimal stubs ~12,000 chars Eliminated boilerplate injection Reduced tool schema via deny list ~5,400 chars Removed session, node, media tools Removed dead remote-node tools (file_fetch, dir_list, dir_fetch, file_write) ~2,580 chars Always fail without a paired node; confused models Fixed t2-list: exec ls instead of dir_list — dir_list requires unconfigured remote node ACL
Total context: ~11,800 → ~9,250 chars per request for local models
Cloud models added to OpenClaw (Anthropic Claude, OpenAI GPT, etc.) automatically receive the full unrestricted tool set via OpenClaw’s tools.byProvider policy. The local provider gets the lean profile; all other providers get browser automation, TTS, image/video/music generation, and all workspace tools. No config changes are needed when adding a cloud API key.