The most common question I get from hardware teams starting an AI product is which chip should we use? The datasheet TOPS number is usually the first thing people compare, and it is the least useful one on its own. This guide walks through the five SoCs I see most often, and how to choose between them based on your model, camera, power budget and supply needs.
The five chips at a glance
| SoC | CPU | NPU | Notes |
|---|---|---|---|
| Rockchip RV1106 | 1× Cortex-A7, 1.2 GHz | 0.5 TOPS (G2) / 1 TOPS (G3), INT8 | ISP for 5 MP at 30 fps, 128–256 MB DDR3L in package |
| Rockchip RK3566 | 4× Cortex-A55 | 0.8 TOPS | Common in low-cost Linux SBCs and tablets |
| Rockchip RK3588 | 4× Cortex-A76 + 4× Cortex-A55 | 6 TOPS, 3 cores | INT4 to TF32 support, 8K video |
| NXP i.MX 8M Plus | 4× Cortex-A53 up to 2 GHz + Cortex-M7 | 2.3 TOPS | Dual ISP, 15-year longevity programme |
| NVIDIA Jetson Orin Nano Super | 6× Cortex-A78AE, 1.7 GHz | 67 TOPS sparse / 33 TOPS dense | 8 GB, 7–25 W modes, $249 dev kit |
Step 1: Start from your model, not the chip
Write down the model you need (for example a YOLO-class detector), the input resolution and the frame rate the product must hit. Then check each vendor's toolchain for operator support: Rockchip's RKNN-Toolkit2, NXP's eIQ, and NVIDIA's TensorRT. Unsupported operators are where most edge AI projects lose weeks.
Step 2: Look at the camera pipeline
For camera products the ISP matters as much as the NPU. The RV1106 and i.MX 8M Plus both include ISPs, so a raw MIPI sensor can be tuned on-chip. Check how many cameras you need, at what resolution, and whether you need hardware video encoding for streaming or recording.
Step 3: Budget power and heat
The RV1106 suits small, low-power cameras. The RK3588 and Jetson Orin Nano need real thermal design: the Orin Nano Super offers power modes up to 25 W for full performance. If the product runs on a battery or in a sealed enclosure, this step often decides the chip on its own.
Step 4: Check software maturity and supply
- Rockchip: strong price and performance, vendor SDKs that usually need work before they are product-ready.
- NXP: mainline-friendly Yocto BSPs and a published 15-year longevity commitment, which matters for industrial and medical products.
- NVIDIA: the most mature AI software stack, at a higher module cost and power draw.
Step 5: Benchmark on real hardware before committing
Convert your actual model, quantise it with calibration images from your real camera, and measure end-to-end latency (capture to result), not just inference time. A week of benchmarking on two or three dev kits is far cheaper than a board respin.
Common mistakes
- Choosing by TOPS alone, then finding half the model runs on the CPU.
- Quantising with internet images instead of frames from the product's own camera, and losing accuracy.
- Ignoring memory: the RV1106 has 128 or 256 MB, which limits model size and frame buffers.
- Picking a high-power module, then discovering the enclosure cannot remove the heat.
Frequently asked questions
Is the RV1106 enough for object detection?
For small, quantised detectors at modest resolution, often yes. It is widely used in smart cameras. Benchmark your exact model first, because 0.5 to 1 TOPS and 128 to 256 MB of RAM leave little margin.
When is a Jetson worth the extra cost?
When you need large models, several high-resolution streams, or the fastest path to a working prototype, and the product can supply the power and cooling.
Can I start on one chip and move to another later?
Yes, if the model pipeline is kept portable (for example ONNX as the exchange format) and the application is separated from vendor SDK calls.
Do I still need an NPU if I have an ESP32-S3?
Not always. Small classification and detection models can run on an ESP32-S3. See my client story on replacing a Raspberry Pi with an ESP32-S3.
Sources
Embedded systems consultant at TechProgeny, Bengaluru. 7+ years and 29+ shipped products across firmware, embedded Linux, edge AI, apps and IoT servers.