Sigmastar Sdk
The newest version of the Sigmastar SDK (v5.x and v6.x) focuses heavily on NPU pipelining. To achieve 30 FPS inference on a 2MP stream, you must avoid copying memory from the ISP to the NPU.
Best practice:
Use the MI_SYS_Bind function to create an in-memory pipeline:
Sensor -> ISP -> VPE -> NPU (NN) -> VENC
This prevents "User Space copying," which is one of the biggest performance bottlenecks in embedded vision. sigmastar sdk
You won't get JTAG on most SigmaStar boards. Here’s your toolkit: The newest version of the Sigmastar SDK (v5
To master the Sigmastar SDK, you must understand its modular architecture. When you download the SDK (typically a compressed .tar.bz2 file), you will find several distinct directories: You won't get JTAG on most SigmaStar boards
SigmaStar heavily utilizes Buildroot. The SDK is not a standard Yocto distribution; it is a heavily modified Buildroot environment. This means:
If you’ve ever worked with a Novatek (now SigmaStar) chipset—like the SSC338Q, SSD202, or SSC30KD—you know the SDK is both a blessing and a beast. It’s incredibly powerful for multimedia processing, but the documentation is often sparse, and the codebase feels like a time capsule from the early 2000s.
After spending months wrestling with the MI (Media Interface) API and the majestic streaming framework, I’ve compiled a list of real-world tips that will save you hours of debugging.