AI Post Transformers / Visual Companion

Caffeine: A Unified FPGA for CNNs

A whole-network FPGA story from the AlexNet-to-VGG era: unify convolution and fully connected layers behind one matrix engine, then win or lose on memory traffic, reuse, and burst scheduling rather than raw multiply counts alone.

Main PDF arXiv trail Memory trade-offs episode

Paper context

Published

2016

Key claim

1 FPGA / whole CNN

Main tension

Compute vs bandwidth

This page uses schematic, normalized values to visualize the paper’s argument: once convolution improves, fully connected layers can become the dominant data-movement bottleneck.

References

Compact trail of papers and adjacent episodes behind the visuals.

Caffeine: A Unified FPGA for CNNs PDF

ImageNet Classification with Deep Convolutional Neural Networks arXiv:1207.0580

Very Deep Convolutional Networks for Large-Scale Image Recognition arXiv:1409.1556

Deep Residual Learning for Image Recognition arXiv:1512.03385

Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep CNNs arXiv:1602.06443

In-Datacenter Performance Analysis of a TPU arXiv:1704.04760

Optimizing FPGA-Based Accelerator Design for Deep CNNs FPGA 2015

AI Post Transformers: FlatAttention for Tile-Based Accelerator Inference