ENH: GPU support on Apple hardware (MPS)

[fransenglich]

Feature Type

• Adding new functionality to pandas

• Changing existing functionality in pandas

• Removing existing functionality in pandas

Problem Description

Pandas got GPU support via cuDF.pandas/Nvidia CUDA, but Apple support is lagging behind here (?). As far as I can tell there’s no effort to write an MPS backend for Pandas, currently. I believe there's a significant amount of users on Apple hardware. I request an MPS backend.

Feature Description

Preferably Pandas detects and use available hardware automatically and appropriately.

Alternative Solutions

Use Pandas as is, but it's comparatively very slow, and doesn't take into account modern hardware.

Additional Context

No response

[rhshadrach]

Thanks for the request. pandas itself provides no GPU support, cuDF is a separate package. I think this is out of our scope.

[fransenglich]

See the ChatGPT answer for “Why is there no MPS/GPU backend for Pandas on Mac?” ChatGPT's answer seems a lot biased towards Nvidia, and rambles if one asks for references.

[jbrockmendel]

Yah I wouldn't even know where to start on this. @seberg might have some ideas. I suspect the first step would be "convince Apple to devote some real resources to making this happen"

[seberg]

I don't really have any input, Apple engineers have helped out with some things in NumPy but not as such to use MPS (I am not quite sure about SIMD, at least for bug-fixes).

cudf.pandas is technically not a backend, I would say. I.e. pandas isn't involved in it at all as such. Rather cudf.pandas does some very advanced stuff to wrap pandas with something that looks and feels the same but uses the GPU for most things.

Would it make sense for pandas to support backends in places? Maybe, I like more backend-style things, but I would think more downstream libraries (dataframe/array users).
My open-sourcy opinion on this is probably that someone should start accelerating core algorithms with MPS separately and then see if the cudf.pandas or even a backend route make sense. But one reason why I like the idea of backends is that the main projects doesn't have to initiate this at all. (And I think that helps all sides, unless e.g. pandas happened to jump at this idea.)

[jbrockmendel]

The word "backend" is pretty overloaded. Can you expand a bit on what it would mean for "pandas to jump at this idea"?

[seberg]

Ah, sorry. I guess there are two paths: one is allowing third party "plugins" for all or specific functions. The other might be to host different versions yourself (like the SIMD kernels in NumPy, I guess).
I meant that I doubt you want to write MPS code inside the pandas repo, so if you don't like the idea (at least long-term) of an mps_pandas monkeypatching pandas, then a plugin system may make sense.
(That is quite different from cudf.pandas, my understanding is that cudf.pandas really needs to do a lot more than replace a few core algorithms.)

[jbrockmendel]

OK. What you're describing sounds a lot like the ExtensionArray design we already have in place. If someone wanted to write an EA subclass that uses GPU memory and compute they could do that today.

[fransenglich]

I don’t work for Apple but have ex-colleagues that now work at Apple. Before filing this report I bounced the topic with one of them, and there’s interest in knowing whether’s there's any capability mismatch that the MPS team isn’t aware of. So there’s likely channels for Pandas/NumPy-folks if hinders are encountered with MPS.

As I see it, if Pandas on Apple hardware is incomparably slower than on Nvidia, Pandas will reduce in interest and other alternatives will become more relevant.

GPU computation isn’t going away any time soon — but the opposite — so the sooner one can do the necessary changes to course, the better, I’d say.