1. TF uses a more basic abstraction in the programming model, i.e., TF.Varaible, than SINGA which is based on the Layer abstraction. Hence their data flow graphs have some differences. 2. SINGA supports distributed training better than the current version of TensorFlow, I think. Both synchronous and asynchronous distributed training frameworks are supported in SINGA. 3. Performance (efficiency, memory cost and scalability) would be compared in v0.3.

It supports multiple GPUs in a single node in this version. Support for GPU cluster is scheduled in the next version referring to http://singa.incubator.apache.org/develop/schedule.html