README
Edge AI Benchmarks — real FPS numbers, no vendor slides
I benchmark China's edge-AI hardware scene: street prices, BOM math, and the FPS numbers vendors would rather you didn't compare. This repo holds the raw data behind RoboKrunch — every figure traceable to a datasheet, a vendor table, or a receipt.
YOLO FPS per dollar (September 2026)
| Platform | NPU | Model | FPS | Board price | FPS/$ | Source |
|---|---|---|---|---|---|---|
| Orange Pi 5 Max (4GB) | RK3588, 6 INT8 TOPS, single NPU core | YOLOv5s | 66.1 | $75 | 0.88 | Rockchip model zoo |
| Orange Pi 5 Max (8GB) | RK3588, 6 INT8 TOPS, single NPU core | YOLOv5s | 66.1 | $95 | 0.70 | Rockchip model zoo |
| Hailo-8 M.2 module | Hailo-8, 26 dense INT8 TOPS | YOLOv5m | 156 | ~$215 | 0.73 | Hailo official tables |
| Jetson Orin Nano 8GB | Orin Nano, 40 sparse INT8 TOPS | — | — | $299 MSRP / $369–$800+ street | — | NVIDIA |
Machine-readable: benchmarks.csv
Read this before you screenshot the table:
- The FPS figures are vendor-reported, not from my lab (yet). Rockchip's own model zoo claims 66.1 FPS on YOLOv5s with a single NPU core; Hailo's tables claim 156 FPS on the larger YOLOv5m. Independent verification is on the roadmap — that's the point of this repo.
- TOPS are not fungible. NVIDIA's 40 are sparse INT8, Hailo's 26 are dense INT8, Rockchip's 6 are INT8 on yet another architecture. Cross-vendor TOPS division is directionally useful and precisely wrong.
- The YOLO variants differ (v5s vs v5m), so the FPS/$ column is a rough price-performance sketch, not a ranking. The curve is smooth, not a cliff: you get what you pay for.
- Prices are September 2026 street/list prices and move constantly. Treat them as directional.
Also measured
- LLM on RK3588 CPU (not NPU): community members have coaxed Qwen2.5-0.5B to ~12 tokens/sec via llama.cpp. A party trick, not a product strategy — the NPU can't run generative workloads, and neither can this CPU at any serious scale.
The analysis behind the numbers
- Full breakdown of where the 4x price gap goes (Shenzhen supply chain, the software tax, BOM math): The $75 AI Computer
- Weekly dispatch on China's edge-AI hardware scene: robokrunch.com
Contribute your numbers
Got a board and a stopwatch? Open a PR or issue with: board, SoC/NPU, model + input size, FPS, how you measured it, and what you paid. Vendor tables welcome too — label them as such and link the source.
Corrections
Found an error? That's what this repo is for. Open an issue or email service@robokrunch.com. Corrections are published, not buried.
License
Code: MIT. Data (benchmarks.csv): CC-BY-4.0 — use it, just credit RoboKrunch.
Contributors
1 commits
See what people are saying
Similar repos
README
Edge AI Benchmarks — real FPS numbers, no vendor slides
I benchmark China's edge-AI hardware scene: street prices, BOM math, and the FPS numbers vendors would rather you didn't compare. This repo holds the raw data behind RoboKrunch — every figure traceable to a datasheet, a vendor table, or a receipt.
YOLO FPS per dollar (September 2026)
| Platform | NPU | Model | FPS | Board price | FPS/$ | Source |
|---|---|---|---|---|---|---|
| Orange Pi 5 Max (4GB) | RK3588, 6 INT8 TOPS, single NPU core | YOLOv5s | 66.1 | $75 | 0.88 | Rockchip model zoo |
| Orange Pi 5 Max (8GB) | RK3588, 6 INT8 TOPS, single NPU core | YOLOv5s | 66.1 | $95 | 0.70 | Rockchip model zoo |
| Hailo-8 M.2 module | Hailo-8, 26 dense INT8 TOPS | YOLOv5m | 156 | ~$215 | 0.73 | Hailo official tables |
| Jetson Orin Nano 8GB | Orin Nano, 40 sparse INT8 TOPS | — | — | $299 MSRP / $369–$800+ street | — | NVIDIA |
Machine-readable: benchmarks.csv
Read this before you screenshot the table:
- The FPS figures are vendor-reported, not from my lab (yet). Rockchip's own model zoo claims 66.1 FPS on YOLOv5s with a single NPU core; Hailo's tables claim 156 FPS on the larger YOLOv5m. Independent verification is on the roadmap — that's the point of this repo.
- TOPS are not fungible. NVIDIA's 40 are sparse INT8, Hailo's 26 are dense INT8, Rockchip's 6 are INT8 on yet another architecture. Cross-vendor TOPS division is directionally useful and precisely wrong.
- The YOLO variants differ (v5s vs v5m), so the FPS/$ column is a rough price-performance sketch, not a ranking. The curve is smooth, not a cliff: you get what you pay for.
- Prices are September 2026 street/list prices and move constantly. Treat them as directional.
Also measured
- LLM on RK3588 CPU (not NPU): community members have coaxed Qwen2.5-0.5B to ~12 tokens/sec via llama.cpp. A party trick, not a product strategy — the NPU can't run generative workloads, and neither can this CPU at any serious scale.
The analysis behind the numbers
- Full breakdown of where the 4x price gap goes (Shenzhen supply chain, the software tax, BOM math): The $75 AI Computer
- Weekly dispatch on China's edge-AI hardware scene: robokrunch.com
Contribute your numbers
Got a board and a stopwatch? Open a PR or issue with: board, SoC/NPU, model + input size, FPS, how you measured it, and what you paid. Vendor tables welcome too — label them as such and link the source.
Corrections
Found an error? That's what this repo is for. Open an issue or email service@robokrunch.com. Corrections are published, not buried.
License
Code: MIT. Data (benchmarks.csv): CC-BY-4.0 — use it, just credit RoboKrunch.
See what people are saying
Similar repos
Contributors
1 commits