Running out of DSP slices on your FPGA? I’ve been experimenting with Low-Rank Approximations for 3x3 Convolutions to solve exactly that.
My latest project replaces standard matrix multiplications with learned, hardware-friendly bit-shifts.
The result:
• 33% reduction in DSP usage (2 Muls instead of 3)
• <1% error (SSIM > 0.99)
• Ideal for SWaP-C constrained edge perception.
https://www.dockerr.blog/blog/lowrank-hardware-approximation
#FPGA #VHDL #EmbeddedSystems #ComputerVision #EdgeAI #OpenScience #Engineering