These are chat archives for dmlc/mxnet

9th
Mar 2016
Sufeng Niu
@sufengniu
Mar 09 2016 21:25
Hello, I am quite fresh here. Currently, I am working on the FPGA based implementation for Deep Learning, not sure whether MXnet community has interested on it or not. but I would appreciate if some one who familiar the MXNet can provide some guidance regarding to the integration with FPGA. and I hope to open source the hardware acceleration to existing popular deep learning repositories.
I find the mshadow is the repository that responding to do the computation engine (Please also correct me if I am wrong). Thus, currently, I am trying to interact the mshadow with the FPGA code
  1. Does the mshadow only offload the matrix multiplication on gpu?
  2. In terms of FPGA, it is not only good at the matrix multiplication, but also very good at the convolution as well (it has specified architecture for convolution). Additionally, the FPGA has inherent structure for pipelining (for example, a 3 layer of feedforward neuron network can be executed simultaneously). In this case, how can I address the interface between mshadow and MXNet, or is the interface provided to the MXNet will be changed?