Which FGPA/tool for this project?

[PartialDischarge]

Latency I can tolerate. However the system must be light, portable, self-contained and battery operated  Sorry I can't disclose the application I've thought of

Hope you're not thinking of using this for a scientific application. Those commercial 3-axis MEMS magnetometers are absolute shite compared to fluxgate or search coil mags.

Ok, please find me a 3D fluxgate sensor that occupies less than 1 or 2 cm3. For this application 3 axes are needed at each point

[DeanoC]

Most of the chips based on mobile soc also have a GPU.

I'd imagine the rpi0w gpu has enough gpu performance to make it worth thinking about. Last time I checked its all ASM based (don't think theres an OpenCL port to the broadcom society) but its 12+ cores, dedicated read caches etc.

The nice thing of course, is that you can write the app for just the CPU on a rpi0w board and then offload calculations to the gpu if you need more speed.

[technix]

Or if you intend to you a more common GPU with proper OpenCL/CUDA you can try a nVidia Jetson kit. This way you can develop on your GTX 1080 Ti to polish the algorithm, then drop the CUDA or OpenCL code onto Jetson’s Maxwell-class GPU.

[technix]

Or if you intend to you a more common GPU with proper OpenCL/CUDA you can try a nVidia Jetson kit. This way you can develop on your GTX 1080 Ti to polish the algorithm, then drop the CUDA or OpenCL code onto Jetson’s Maxwell-class GPU.

The original question was to get a processor that is 5x~10x as fast as a 67MHz CM3, so the ideal choice should be a mid-low range DSP or CA5/CM7. Now it's gone out of control to TFLOPS range .

I don’t think Jetson is getting out of hands, and it is Hetson that is going into the project. The VC4 used in Pi isn’t that programmable without Broadcom unobtainium tools. Jetson packs a few Maxwell-class GPU cores that runs the easily accessible CUDA and OpenCL tools.

The GTX 1080 Ti runs the same CUDA and OpenCL. Hence it is used for faster development and algorithm optimization.

[technix]

I don’t think Jetson is getting out of hands, and it is Hetson that is going into the project. Jetson packs a few Maxwell-class GPU cores that runs the easily accessible CUDA and OpenCL tools.

A Jetson still packs somewhere around 1TFLOPS while the OP needs 1GFLOPS or 1GIPS. Why wasting $3-digit on 1000x needed performance?

Depending on the real time requirement, you cannot load the system beyond a certain point without losing responsiveness.

Also did I mention tools? Good luck getting Broadcom tools for VideoCore IV.