You need to have a Titan X to produce comparable benchmarks, otherwise a lot of people won't take them seriously. Also the 6GB of a 980Ti already rule out quite a few popular networks we needed to run.
I was not really in a position to wait much when buying the card, and I didn't want to risk waiting a few weeks to get a used card I bought on a forum, which then comes out being just as expensive in living cost. You also as a startup don't generally have the luxury to risk waiting.
I don't regret the purchase, I more wanted to show the reality of a "hard tech" startup.
You mean on AWS/Azure? Because I have to develop a framework (https://github.com/autumnai/leaf) with it, and there is no way I am going to introduce any more complexity than necessary that early.
Understood. Leaf looks interesting. Will be keeping an eye on it. Having just recently focused on machine learning I still haven't settled on a stack. I find your rational behind using Rust quite helpful. What kind of timeframe are we looking at for when this is production ready?
Really depends on what use case. There should be a nice production-ish demo out towards end of the month, but I am afraid I can't say anything less vague at the moment.
Does the rust core team have any plans towards being able to write kernels directly in Rust using the NVPTX LLVM backend? Otherwise if you're all just calling into CuDNN anyway, pretty narrow differentiation in performance (and big difference in development resources) between you, Torch, and TensorFlow.
Source: I just did that.