Alternative Computing Technologies (ACT) Lab

School of Computer Science

Georgia Institute of Technology

DnnWeaver is the first open-source framework for accelerating Deep Neural Networks (DNNs) on FPGAs. While FPGAs are an attractive choice for accelerating DNNs, programming an FPGA is difficult. With DnnWeaver, our aim is to bridge the semantic gap between the high-level specifications of DNN models used by programmers and FPGA acceleration. With our framework, the programmer only specifies the Deep Neural Network using Caffe format. The framework automatically generates the accelerator Verilog code specialized for the given network, using our hand-optimized Verilog templates. DnnWeaver is under development at the Alternative Computing Technologies (ACT) Laboratory, Georgia Institute of Technology. Below, you can download our framework and the Verilog code for our template architecture.


Citation

If you use this work, kindly cite our paper (PDF, Bibtex) published in The 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), 2016.

H. Sharma, J. Park, D. Mahajan, E. Amaro, J. K. Kim, C. Shao, A. Mishra, H. Esmaeilzadeh, "From High-Level Deep Neural Models to FPGAs", in the Proceedings of the 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), 2016.

Workflow



DnnWeaver consists of four components:

Translator:

The Translator component of DnnWeaver creates a macro-dataflow graph of the DNN model using our Instruction Set Architecture (ISA). Each node of the macro-dataflow graph represents a layer of DNN model. DnnWeaver then generates a static execution schedule for the accelerator using the macro-dataflow graph.

Design Planner:

The Design Planner optimizes the following parameters to specialize the accelerator for the given DNN model:

  1. The organization of the accelerator to maximize performance.
  2. The slicing of the data to minimize off-chip memory accesses.
The organization of the accelerator depends on the available compute resources on the FPGA (DSPs and LUTs) and the slicing depends on the available on-chip memory (BRAMs).

Design Weaver:

The Design Weaver generates the accelerator core using our hand-optimized Verilog templates according to the Design Planner.

Integrator:

The Integrator is the final component of DnnWeaver and appends the memory interface code to the accelerator. As a part of the initial release, DnnWeaver includes the memory interface code for Xilinx Zynq ZC7020 board.


Benchmark List

DnnWeaver supports a wide range of Convolutional Neural Networks. The following are the benchmark models we used for the experimental results in our paper:

Benchmark Deep Neural Network (DNN) Models
# Benchmark Name Data Set Description Number of Layers Model Size (MegaBytes) Number of Operations Lines of Code (prototxt)
1 LeNet MNIST Hand-written Digit Recognition 7 0.8 MB 2 GOps 128
2 Cifar-10 Full Cifar-10 Object Recognition 12 0.2 MB 12 GOps 156
3 Network-in-Network (NiN) ILSVRC 2012 Object Detection and Classification 28 14.5 MB 1106 GOps 516
4 Djinn-ASR Djinn and Tonic Speech-to-text Decoder 13 48.4 MB 25 GOps 105
5 AlexNet ILSVRC 2012 Object Detection and Classification 20 119.0 MB 1147 GOps 278
6 VGG-CNN-S ILSVRC 2012 Object Detection and Classification 19 196.0 MB 2666 GOps 200
7 Overfeat ILSVRC 2012 Object Detection and Classification 16 278.0 MB 2798 GOps 196
8 VGG-16 ILSVRC 2012 Object Detection and Classification 36 324.0 MB 16362 GOps 347

License

Copyright 2017 Hadi Esmaeilzadeh
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

      http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Developers

Hardik Sharma, Jongse Park, Divya Mahajan, Emmanuel Amaro, Joon Kyung Kim, and Hadi Esmaeilzadeh


Maintained by

Hardik Sharma


Contact

Please forward all your inquiries to: hsharma@gatech.edu


Patrons