Engineer adouglass52 designed a custom compute module in the Raspberry Pi CM4 format, built around the Rockchip RV1126B, a vision processor with a 3 TOPS NPU. With the first boards just back from JLC, the maker ran it against a CM4 (CM4101000) on the same carrier board with the same two IMX415 cameras, using stock tools on both and the median of three runs.
What is on the board
The module pairs the RV1126B (four Cortex-A53 cores at 1.2 GHz) with 1 GB of LPDDR4X, 8 GB of eMMC, 16 MB of SPI flash and Wi-Fi 6. It follows the CM4 pinout closely enough to drop onto a CM4 carrier. It is not fully pin-compatible, though: the second 4-lane camera borrows the DSI0 pins, so on a stock CM4 carrier that camera runs as 2-lane. The maker's cost to build at 1,000 units at JLC is about 62 dollars before shipping and tariffs, against a 71.25 dollar list price for the comparable CM4101008.
The numbers
On vision work the custom board wins clearly. It ran YOLOv8n at 640x640 on its NPU in INT8 at 25.4 ms per frame, or 38.9 fps. The CM4 has no NPU and managed 4.0 fps on its CPU, about 10x slower. The RV1126B board also captured two IMX415 cameras at 4K30 at the same time, which the CM4 could not do, encoded 1080p camera video to H.264 at 30 fps against the CM4's 25 fps, and added H.265 encoding, which the CM4 lacks. Encoding 1080p from RAM ran at 95 fps against 72, and it reached 26 fps at 4K, where the CM4 has no 4K encode.
The CM4 still wins on raw CPU: about 2x in a 4-thread sysbench run (2,315 versus 1,160 events per second) and ahead in 7-Zip (6,520 versus 5,456 MIPS). It also keeps HDMI and PCIe, which the RV1126B module does not have. The custom board drew about 0.5 A at full CPU load against about 0.9 A for the CM4, ran about 20 degrees C cooler, and got roughly twice the Wi-Fi throughput on the same access point and antenna.
Our take
This is a rare apples-to-apples comparison of a CM4 against a purpose-built vision chip, from someone who had to bring up the hardware and firmware from scratch. The takeaway is simple: for cameras, video and on-device object detection, a small NPU beats a faster CPU by a wide margin. If you are new to vision on embedded boards, our ESP32-CAM object detection project is an easy place to start.
Image credit: adouglass52 / Reddit, cropped and resized.
About the author: Aswinth Raj is an electronics engineer and the editor of CircuitDigest, where he has published more than 150 hands-on projects and tutorials across Arduino, PIC, ESP32, Raspberry Pi and EV battery systems. He selects the stories on this page for what engineers and makers can actually learn or build from. Connect on LinkedIn
Add a comment
No account needed. Your comment appears once a moderator has reviewed it.