Spectral-Blaze: A High-Performance FFT-Based CNN Accelerator

This paper presents Spectral-Blaze, a novel FFT-based CNN accelerator that effectively addresses computational and energy bottlenecks in spatial domain acceleration. The proposed architecture introduces Intra-Patch parallelization during the Hadamard product phase, which optimizes Complex MAC (CMAC) unit utilization and maintains consistent reuse patterns across multiple input feature map patches. This parallelization scheme simplifies the tiling, enabling independent selection of patch elements for efficient on-chip memory storage. Additionally, it leverages the Flex-Stationary dataflow to adaptively store tensors with high reuse opportunities in the on-chip memory, reducing the DRAM traffic. Spectral-Blaze also introduces an optimized FFT/IFFT process that integrates a modified row-column FFT, real FFT-based optimizations, and a vector-radix IFFT. The Spectral-Blaze prototype, implemented on the Zynq MPSoC (ZU7CG), demonstrates impressive performance gains, achieving a speedup of \(4.98\times \) for VGG-16 and \(1.64\times \) for AlexNet compared to baseline. The Energy-Delay Product (EDP) varies from \(4.47\times \) to \(25.14\times \) . Compared to flexible dataflow spatial accelerators, Spectral-Blaze achieves an average improvement of \(8.82\times \) over Eyeriss and \(7.87\times \) improvement over Flexflow in EDP while maintaining a comparable hardware setup. Spectral-Blaze also achieves a \(4.59\times \) reduction in MAC operations during convolution, even with VGG-16’s compact filter sizes.