Vathys is hiring a machine learning compiler engineer to build the core compiler to map workloads to our novel 3D dataflow processor.
Superficially, a machine learning compiler is just another compiler, but in truth, it is different in a number of important ways:
1. Machine learning compilers must incorporate "statistical" optimizations in addition to "formal" optimizations. For example, an ML compiler might decide to choose a different batch size or a different sparsity level.
2. Optimizations in ML workloads usually have strong "priors". For example, convolutional networks heavily re-use similar convolution kernels and layer connection patterns (e.g. residual blocks).
The ideal candidate has a good understanding of the fundamentals of deep learning algorithms. Less "I can use Keras" and more "I can write a CNN from scratch in numpy".
Compiler familiarity is good of course, but extensive experience in compiler development is not necessary.
Generous compensation will be provided to the right candidates.
Please contact tapabrata_ghosh[at]vathys[dot]ai if you're interested.