Benchmark Radxa Orion O6N with Qwen3.5 35B

devices/radxa-orion-o6n-32gb.md

@@ -0,0 +1,99 @@
+---
+# ============================================================
+# LLM Local Inference Device Benchmark - Data Template
+# Copy this file and rename to: vendor-model.md (lowercase, hyphens)
+# Example: nvidia-jetson-orin-nano-8gb.md
+# ============================================================
+
+# Basic info (required)
+id: "radxa-orion-o6n-32gb"              # Unique ID, same as filename without .md
+name: "Radxa Orion O6N"                 # e.g. "NVIDIA Jetson Orin Nano 8GB"
+vendor: "Radxa"                         # e.g. "NVIDIA"
+device_type: "Dev Board"                # Dev Board / GPU / Accelerator / Mini PC / Server
+
+# Chip info (optional)
+chip: "CIX P1 CD8180"                   # Main chip model
+
+# Memory (required)
+memory_capacity_gb: 32                  # Memory capacity in GB
+memory_bandwidth_gbs: 96                # Memory bandwidth in GB/s
+
+# Memory details (optional)
+memory_type: "LPDDR5"                   # LPDDR4 / LPDDR5 / LPDDR5X / HBM3 etc.
+memory_speed: "6000MT/s"                # e.g. 8533MT/s
+memory_bus_width: "128bit"              # e.g. 256bit
+
+# Claimed compute (optional — due to huge variance in vendor methodologies)
+tops_int8: 30                                   # Claimed TOPS
+tops_note: "NPU (official dokumentation 28,8)"  # Methodology: "GPU only" / "sparse" / "GPU+DLA" etc.
+
+# Price & power (required)
+price_usd: 199                          # Reference price in USD
+power_watts: 31                         # Power consumption under load, in W
+
+# Interface (optional)
+interface: ""                           # PCIe x16 / USB-C / M.2 / Onboard etc.
+
+# ============================================================
+# LLM Benchmark Data (required — at least one Qwen3.5 model)
+# ============================================================
+benchmarks:
+  - model: "Qwen3.5-35B-A3B"            # Model name
+    quant: "int4"                       # Quantization: int4 / fp4 / int8 / fp8 / bf16 / f32
+    framework: "ik_llama"               # Tool: ollama / llama.cpp / LM Studio / vendor SDK etc.
+    decode_tps: "16.13"                 # Output speed in tokens/s (required)
+    prefill_tps: "93.51"                # Prefill speed in tokens/s (optional)
+    context_length:                     # Context length used (optional)
+    image_encode_ms:                    # Image encode time in ms (optional, vision models only)
+
+#  - model: "Qwen3.5-27B"
+#    quant: "int4"
+#    framework: "ollama"
+#    decode_tps:
+
+#  - model: "Qwen3.5-35B-A3B"          # Optional
+#  - model: "Qwen3.5-122B-A10B"        # Optional
+#  - model: "Qwen3.5-397B-A17B"        # Optional
+
+# Submission info (required)
+submitted_by: "Schnabulator"
+date: "2026-04-04"
+---
+
+# Device Name — Benchmark Report
+
+## Test Environment
+
+- **OS**: SKY1 Linux (6.19.4-sky1-latest.r2)
+- **Framework**: ik_llama build: 67fc9c5e (4391)
+- **Model source**: https://huggingface.co/bartowski/Qwen_Qwen3.5-35B-A3B-GGUF (e.g. Hugging Face / Ollama library)
+- **Cooling**: stock (stock / aftermarket / ambient temp)
+- **Power supply**: 65W
+
+## Results
+
+(Additional details, multi-run results, different settings, etc.)
+
+## Evidence
+
+Place screenshots in `devices/assets/<device-id>/` directory.
+For example, if your device id is `nvidia-rtx-4090-24gb`, put images in:
+`devices/assets/nvidia-rtx-4090-24gb/`
+
+Then reference them in this file with relative paths:
+
+![Qwen3.5-35B-A3B benchmark](assets/radxia-orion-o6n-32gb/orion-o6n-bench.png)
+
+Accepted formats: `.jpg`, `.png`, `.webp`
+
+## Device Photo
+
+(Optional: photo of the actual device, placed in the same evidence directory)
+
+![Device](assets/radxia-orion-o6n-32gb/device_photo.jpg)
+
+## Notes
+
+(Any additional information, known issues, special configurations, etc.)
+measured_read_bandwidth: "~109 GB/s (8 threads, sysbench)"
+measured_write_bandwidth: "~94,5 GB/s (8 threads, sysbench)"