RecEasyNet

In this project, a large recommendation model (three-layer forward neural network) is deployed in FPGA clusters. Communication among nodes is realized via Send()/Recv() functions from EasyNet.

Five bitstreams are generated:

1. node_1_mem_first: as the start node, do memory reading and computation in layer 1
2. node_1_compute: do computation only in layer 1
3. node_1_mem: do memory reading and computation in layer 1
4. node_2: do computation in layer 2
5. node_3: do computation in layer 3 and output layer

The basic communication pattern is in a ring method like: node_1_mem_first -> node_1_compute -> node_1_mem -> node_1_compute ... -> node_1_mem -> node_1_compute -> node_2 -> node_3.