[please boost for reach]
I'm looking for a single-board computer with some specific features:
- Two MIPI-CSI camera ports
- Decent CPU/GPU, on par with RasPi 4 performance
- An accelerated, open-source graphics stack (*)
The only option I see right now is to use a CM4 with a custom carrier board - any alternatives out there?
(*) I'm aware that this is really difficult to get without any binary blobs at all, but please, at least no outdated custom Android images without build instructions 😑
@floe Not going to have luck with that on Arm SoCs. Nvidia has a couple of expensive SBCs with 2x CSI, but those need blobs.
Rockchip has some interesting chips like the RK3588, but imho it's too early for prod use.
You also haven't mentioned yet what you need the GPU/graphics stack for.
Configuring the ISPs/GPUs to do any real work with camera data is going to be a _REAL_ challenge.
@manawyrm I'm exploring options for an open-source augmented reality headset. For starters, I'd be fine with doing camera data processing on the CPU and "just" use the GPU for 3D rendering.
I did fiddle around with a RK3399 board a few years ago; AFAICR the Linux support was atrocious. Has Rockchip recently improved in that area?
@floe Well, Rockchip is actually pretty much the best vendor in terms of Linux support (that says a lot I'd say...)
Apart from them, there's pretty much only NXP with their i.MX line, but their graphics implementation is ... eh... well, let's not speak about that.
Yeah, Rockchip might just work for you, with the Panfrost driver you can get semi-decent OpenGL ES performance.
@floe tbh, your application sounds like you want to bite into the sour apple and use the proprietary/blob nvidia driver.
either that or an AMD x86 board and some other solution for the camera data input.
@floe I feel like I'm a bit negative here, sorry about that, the situation really is a bit sad.
Rockchip boards newer than the RK3399 aren't really well supported. RK3588 is worth a look, there's some support there, but it's very very recent (will probably need self-built kernel, mesa, etc.)
@manawyrm I've been digging a bit more (also posting this here for future reference), and the most promising option seems to be something like the OKdo Nano C100: https://www.okdo.com/p/okdo-nano-c100-developer-kit-powered-by-nvidia-jetson-nano-module/
AFAICT you can plug in an Nvidia Jetson (the default), the new Turing RK1 with RK3588 SoC (https://www.cnx-software.com/2023/07/28/turing-rk1-rk3588-system-on-module-for-cluster-boards/), or even a Raspi CM4 via adapter board.
Overall, that seems to be a promising option. Pricey, but I'll perhaps give it a try.
The OKdo Nano C100 Developer Kit powered by NVIDIA® Jetson Nano Module delivers the compute performance to run modern AI workloads at unprecedented size, power, and cost. Developers, learners, and makers can now run AI frameworks and models for applications such as image classification, object detection, segmentation, and speech processing.
Florian 'floe' Echtler
Manawyrm | Sarah