Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Need More GPU for a medical applicaiton
- Thread starter anjanesh
- Start date
-
- Tags
- gpu
- Sort by reaction score
Hmm, there are models available for very task, my team tested some with data from https://nihcc.app.box.com/v/ChestXray-NIHCC on an el cheapo PC cobbled from leftover parts and two retail RTX4090s and I think they went sub-1s/image. I believe this is the most cost-effective solution.
despite how amazing and efficient apple silicon gpu in the top tier m2 ultra, it still no match for even last gen nvidia 3000 series gpu.
you would need to ditch the apple ecosystem in order to get what you asking for.
Forgive the obvious, but have them double-check they are actually using the GPU -- I've seen quite a few cases where people have installed the standard Python tools and just assumed it was using the AS's GPU. You have to make sure you have the right libraries installed.
Beyond that, check out Paperspace. In fact, they have a nice price comparison tool of the various GPU hosting services:
https://www.paperspace.com/gpu-cloud-comparison
For quick and easy, you can fire up colab.google.com and purchase GPU units on an ad hoc basis.
Beyond that, check out Paperspace. In fact, they have a nice price comparison tool of the various GPU hosting services:
https://www.paperspace.com/gpu-cloud-comparison
For quick and easy, you can fire up colab.google.com and purchase GPU units on an ad hoc basis.
In addition to all the above (make sure the code actually is using the GPU), do you know what model MacBook Pro the example is from? For all we know, it could be an Intel machine with integrated GPU from 2015.
"In addition to all the above (make sure the code actually is using the GPU)" - Is there a way to verify this ?
See:
https://developer.apple.com/metal/pytorch/
There is Python code at the bottom of page to check if the GPU is being used.
Also see:
https://pytorch.org/docs/master/notes/mps.html
LambdaLabs is the best bet between cost effectiveness and ease of use. You can do 8xv100 for $4.40 an hour.
Would probably do your 8 hour job in 10 minutes or so, possibly a lot less since you didn't specify which MBP. AWS is workable as well but will be 2-5x the cost depending on what solution you pick.
And the 40 series absolutely destroys 30 series GPUS (except the 4060 - that's a dog)
This reads like a student task. They should be able to get some free credits for MS Azure. Spin up a powerful Windows machine there to run the program, then remove it from the account once done.
eGPU
Use an eGPU with your Mac - Apple Support
Learn how to access additional graphics performance by connecting to an external graphics processor.
Make sure GPU is the actual bottleneck.
Make sure Python is vectorized.
Unfortunately, no eGPU support for Apple Silicon machines as of yet. Probably better to get a Windows machine with a GPU and remote into it like a server anyway, freeing the MacBook from participating.eGPU
Use an eGPU with your Mac - Apple Support
Learn how to access additional graphics performance by connecting to an external graphics processor.support.apple.com
Last edited: