Yes, I'm cofounder and I pretty much live and breathe the company. I see how my comment reads as soulless shilling so I'll lay out my perspective and you can make of it what you will. This is all my personal opinion and not necessarily related to our product or company.

At a basic level I think making new powerful technology accessible to more people is on average strongly positive. There are various efforts making good progress to address different parts of deep learning accessibility such as Keras (developer-friendly Python API), OpenAI (open basic research & safe AI), fast.ai (practical training for developers), etc. I'm a fan of all of that work. PlaidML is the company's contribution to making adoption easier.

For the purposes of proliferation and democratization making deep learning work on the most readily available hardware helps people get started with less friction. PlaidML is a step in that direction. It's fully open source and you can right now 'pip install' it on Mac/Win/Linux with Intel/AMD/NVIDIA GPU and have a Keras net running in a couple minutes. There are certainly warts and some missing features but as far as I know it's the only one an ordinary practitioner can use right now.

From a "what problem does this solve" standpoint PlaidML is most similar to Tensor Comprehensions and TVM. Each makes different tradeoffs but might eventually be able to share components like code generation for OpenCL, LLVM, etc. Layers like XLA, nGraph, ONNX, NNVM, etc, you can mostly think of as being stacked on top (they are ways to talk to lower layer runtimes like PlaidML). For example it would be reasonable for a future version of PlaidML to support TensorFlow integration via XLA or deployment of ONNX models on OpenCL-capable GPUs.

Anyway, I personally care most about what people can use. There's a cute demo that will run the pre-trained Keras examples against images from your webcam on your local GPU. It's quick to try and can serve as the basis for prototyping a real application: https://github.com/plaidml/plaidvision

One of the core operations of the transformer network[1] is a (LxL) x (LxE) matrix multiply (where L is the sentence length and E is the network width). Can you be more specific about how you would get good performance without specializing on L?

[1] https://arxiv.org/abs/1706.03762

The kind of optimizations a static compiler might apply can be done by a JIT as well, with the added benefit of actually knowing what kind of workload is going to run. Most of deep learning is applying comparatively small computation graphs to very large arrays of numbers in parallel, so the overhead of compilation is only a small portion of the overall computation time. A smart JIT that decides on the optimal tiling pattern for the array dimensions observed at runtime and rewrites loops accordingly can easily pay for itself.

That's "course-grained reconfigurable architecture", for anyone else who didn't know.