What it does

PipeCNN compiles convolutional neural network inference kernels to FPGA bitstreams using OpenCL and High-Level Synthesis. You write C-like OpenCL kernels; Intel’s SDK or Xilinx Vitis turns them into RTL, then into a running FPGA design. It ships with pre-quantized VGG-16 and ResNet-50 models and a ModelZoo of weights and test vectors.

The interesting bit

The authors openly state the project is “no longer comparable to the state-of-the-art designs” after four years of DLA evolution. Rather than chase benchmarks, they’ve pivoted to educational value: a complete, working HLS pipeline you can actually modify. That’s unusual honesty in accelerator research, where papers typically pretend last year’s work is still competitive.

Key highlights

• Supports both Intel OpenCL SDK Pro 20.1 and Xilinx Vitis 2020.1
• Tested on Arria-10, Zynq, and Alveo U50 boards; may work on DE10-nano and Ultra96-v2 (unverified)
• Pipelined CNN kernels with tunable parameters (VEC_SIZE, LANE_NUM, CONV_GP_SIZE_X) for per-board optimization
• Includes ImageNet classification demo with OpenCV integration
• Spawned follow-up research including a sparse-convolution variant (DAC 2019)

Caveats

• Performance table in the README is blank (all “–”), so you get no hard numbers for comparison
• Requires specific, somewhat dated tool versions; newer releases may need massaging
• The authors themselves note the design is behind current state-of-the-art

Verdict

Grab this if you’re learning FPGA-based deep learning acceleration or need a hackable HLS baseline to test new ideas. Skip it if you need production throughput today—modern DLA IP or even a GPU will run circles around it.