Translation and accelerated solving of differential equations on GPU platforms
1–10 of 34 posts
Re: Translation and accelerated solving of differential equations on GPU platforms
#2"Instead of relying on high level array intrinsics that machine learning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead." Chris Rackauckas on LinkedIn earlier today.
Re: Translation and accelerated solving of differential equations on GPU platforms
#3"Instead of relying on high level array intrinsics that machine learning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead." Chris Rackauckas on LinkedIn earlier today.
What is direct kernel generation?
Re: Translation and accelerated solving of differential equations on GPU platforms
#4"Instead of relying on high level array intrinsics that machine learning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead." Chris Rackauckas on LinkedIn earlier today.
What is direct kernel generation?
Maybe something like this?
https://www.intel.com/content/www/us/en/developer/articles/t...
Re: Translation and accelerated solving of differential equations on GPU platforms
#5"Instead of relying on high level array intrinsics that machine learning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead." Chris Rackauckas on LinkedIn earlier today.
CR is a hero. The work he does on ODE and related libs in Julia is one of the selling pts of the language. He is the Alex Crichton of Julia.
Re: Translation and accelerated solving of differential equations on GPU platforms
#6Link to GitHub repo from the abstract: https://github.com/SciML/DiffEqGPU.jl
Re: Translation and accelerated solving of differential equations on GPU platforms
#7Earlier quoted context omitted.
What is direct kernel generation?
Maybe something like this? https://www.intel.com/content/www/us/en/developer/articles/t...
Yeah you'd be surprised what the performance gain is for hand written kernels.
There's probably a ton left on the table if you really want to go fast.
Re: Translation and accelerated solving of differential equations on GPU platforms
#8Uhh they time the vmap of the jit on Jax, basically skipping a ton of optimizations,.esp if there is any linear algebra in there. They also include the cost of building the vmap functional. Not a valid comparison.
https://github.com/utkarsh530/GPUODEBenchmarks/blob/ef807198...
Re: Translation and accelerated solving of differential equations on GPU platforms
#9"Instead of relying on high level array intrinsics that machine learning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead." Chris Rackauckas on LinkedIn earlier today.
What is direct kernel generation?
In this context I would imagine it's constructing source code for a kernel- the engine that implements a step in a neural network- that is closer to optimal. See https://cuda.juliagpu.org/stable/tutorials/performance/ for related work
Re: Translation and accelerated solving of differential equations on GPU platforms
#10Uhh they time the vmap of the jit on Jax, basically skipping a ton of optimizations,.esp if there is any linear algebra in there. They also include the cost of building the vmap functional. Not a valid comparison. https://github.com/utkarsh530/GPUODEBenchmarks/blob/ef807198...
What they should do is build the vmap and jit that, then run timing on calling the resulting function.