Python theano dot product too slow

I've been working on a neural network for a while now. I'm currently held up by a bottleneck from dot product computations I think. There's 5000 time steps for 4 different layers of sizes 3, 3600, 900, and 121 so the matrices are pretty big. I'm using mini batching with batch size 10. I'm using numpy and have tried using theano in order to utilize the GPU. I'm not remotely good with this low level stuff so I'm just assuming that the major bottleneck is IO between GPU and CPU. Profiling the code shows that theano's call function (called every time I try to use a GPU-enabled dot product) takes about .008 seconds (.010 if I use float32). I don't know if this is the best speed I can expect due to sheer size of the matrices or if I'm messing up. My code is about 300 lines long so I'll try to post specific snippets if anyone requests.