I am working on a requirement from a Customer who wants to setup a ADAS Platform.

“1. Do the Intel FPGA support Caffe<https://www.zdnet.com/article/caffe2-deep-learning-wide-ambitions-flexibility-scalability-and-advocacy/>, CNTK<https://www.microsoft.com/en-us/cognitive-toolkit/>, DeepLearning4j<https://www.zdnet.com/article/machine-learning-in-the-cloud-is-the-new-battlefield/>, H2O<https://www.h2o.ai/>, MXnet<https://mxnet.apache.org/>, PyTorch<https://pytorch.org/>, SciKit<http://scikit-learn.org/stable/>, and TensorFlow ? Which model of FPGA?

Our OpenVINO tool (Platform agnostic) supports Caffe, TensorFLow, MxNet…for sure. Refer to OpenVINO webpage for more info

2. Are FPGA suggested for use in Deep Learning Training? Which model ?

FPGAs can do training and we have used them for the forward (Inferencing) path, but we encourage use of Nirvana Crest family of devices for training acceleration. FPGAs focus on inference really.

3. How do they compare to NVIDIA Tesla V 100?

Depends on model, batch size, power, TCO of the system. IN general the GPU can be more images/sec, but the question is about deployment in a system. WE keep the data in the FPGA and between FPGAs without needing to go back and forth to the host, so we often have higher system performance and almost always significantly lower latency.

4. Can FPGA be stacked, to get the required throughput?

Yes. This is especially beneficial when persistence is achieved. FPGAs tend to get significant performance gains here.

5. Which one to use and where ? Intel Nervana –T or Intel Arria 10 or Intel Stratix 10?”

Nervana is for training. Arria 10 is the only FPGA inference solution we have right now. WE will be supporting S10 shortly when D5005 card is officially released and OpenVINO supports it