I might be egged in this forum, or tarred & feathered, but my suggestion is to switch to a different platform: NVidia Jetson Orin.
This is much more powerful in terms of computation power. You get more cores, you get a fat, CUDA-capable GPU with Dual 4k@60 capable DP++/HDMI ports. Plus there is a dedicated Cortex R52 Real-time Sensor Processor core thrown in. You get much better I/O. Ok, UARTs, SPI, I2C is there as well. But you also have one PCIe Gen 4 x4, one PCIe Gen 4 x1, and either one PCIe Gen 4x2 or two PCIe Gen 4 x1 busses on the small DIMM-based platforms. High-bandwidth connections are so much easier with good PCIe connectivity. Streaming of high-resolution compressed video over 2.5 or 10G Ethernet (Twisted Pair or Fiber) becomes easy when you have a powerful GPU that handles compression/decompression.
I've designed systems that read video streams from a 50Mpix full format camera connected via PCIe 2.0 x4 directly from the camera chip into the shared main memory, processes it, and yanks it out onto an 8TB U.2 NVME SSD.
Of course the processor modules are more expensive. You can choose between different CPU/GPU core counts and memory sizes, so you can upgrade easily without throwing everything away.
BTW: I like the Compute Modules for various reasons. But in computing there is no one-size-fits-all.
This is much more powerful in terms of computation power. You get more cores, you get a fat, CUDA-capable GPU with Dual 4k@60 capable DP++/HDMI ports. Plus there is a dedicated Cortex R52 Real-time Sensor Processor core thrown in. You get much better I/O. Ok, UARTs, SPI, I2C is there as well. But you also have one PCIe Gen 4 x4, one PCIe Gen 4 x1, and either one PCIe Gen 4x2 or two PCIe Gen 4 x1 busses on the small DIMM-based platforms. High-bandwidth connections are so much easier with good PCIe connectivity. Streaming of high-resolution compressed video over 2.5 or 10G Ethernet (Twisted Pair or Fiber) becomes easy when you have a powerful GPU that handles compression/decompression.
I've designed systems that read video streams from a 50Mpix full format camera connected via PCIe 2.0 x4 directly from the camera chip into the shared main memory, processes it, and yanks it out onto an 8TB U.2 NVME SSD.
Of course the processor modules are more expensive. You can choose between different CPU/GPU core counts and memory sizes, so you can upgrade easily without throwing everything away.
BTW: I like the Compute Modules for various reasons. But in computing there is no one-size-fits-all.
Statistics: Posted by fchk — Tue Apr 29, 2025 6:36 pm