https://github.com/Pascal-J/Jfire
Lush: it just felt comfy for a data science workflow. No DB adapters, or even real csv readers, but you could do whatever you needed in the interpreted language and go pretty fast with the compiled bits. Still feels like the future.
I do wish we had some better cross-platform solutions though, instead of cementing nVidia's hegemony. There's no technical reason why GPGPU should be any more difficult on AMD hardware.
Now, all vendors seem to have dropped support for OpenCL, though it will live on nicely in PortableCL, which implements it atop LLVM, allowing it to target anything LLVM does, which finally goes further to achieve the open-ness the original standard aspired to, than the vendors themselves at the time.
Then look no further than GPUCompiler.jl. Tim factored out all the platform agnostic parts of CUDA.jl (successor of CuArrays.jl and CudaNative.jl) into GPUCompiler.jl and that is now being used by the AMDGPU.jl package. It works today!
That just takes developer time to fix. Currently it’s basically just a one-man show.
Recently I have been trying to learn Futhark which is a ML like language generating parallel code (E.g. OpenCL or CUDA) from pure functional code with nested array structures. I find this exciting because when using the GPU with a thin layer like OpenCL directly, you have to write kernels from scratch without much abstraction possible. Higher level APIs like PyTorch or Jax give you easy ways to flexibly structure the computation but you just have to hope for kernel fusion. In Futhark this is in principle explicit and well handled.
ISPC is another example of a C like language where SIMD style parallelism is unambiguous while staying relatively high level.