I’m just starting 🔥 Introduction - Mojo 🔥 GPU Puzzles, and on Puzzle 01 am getting the error shown below. I assume this is because my laptop’s GPU is not yet supported: GP107M [GeForce GTX 1050 Mobile]. If this is the case, can I somehow configure my setup to use cloud compute for GPU? Thanks in advance for any help/guidance! Error
open-source/max/mojo/stdlib/stdlib/builtin/_startup.mojo:95:4: error: call expansion failed
open-source/max/mojo/stdlib/stdlib/builtin/_startup.mojo:65:4: note: function instantiation failed
/home/johnsoe1/dev/Modular/github/mojo-gpu-puzzles/solutions_eaj/p01/p01.mojo:25:23: note: call expansion failed
with DeviceContext() as ctx:
Mojo and MAX won’t magically use cloud compute. The easiest way is to use the vscode “SSH” plugin to connect to the lambda cloud GPU instance and install Mojo on there.
Your GPU isn’t supported right now, and it will take a bit to do the work to make it function since it’s the one of the older GPUs that would work at all with current kernels on the Nvidia side. Your best bet is to use cloud compute.
I had a look into supporting the Pascal architecture after the community success with Turing, but it looks like a fair bit of work.
I added GP104 (GTX 1070) to the build scripts and added extra constraints around asynch and tensorcore features which are not available on Pascal. It builds but a number of the tests fail due to the JIT emitting code that doesn’t work on this architecture. So some work on the backend would be needed. Since Pascal is already EOL for Nvidia the buck probably stops here
Thanks for looking at it @apnea! I realized during the hackathon that not only was a Lambda instance not a good long-term solution for GPU development, but my laptop with the GTX 1050 is also seriously constrained by memory (16 GB) and hard drive space. I have since replaced it with a new Lenovo laptop with a Blackwell RTX 5090 GPU, so I now have a decent local setup for GPU development [stoked!! ].
Pascal is not technically EOL yet, and it will take a long time before there is no supported driver due to Nvidia’s need to support older DXG systems. On windows, it will probably go away fairly quickly, but on Linux we should be able to keep it around for years. There are a lot of boxes of 1080tis which still have enough vram to run modern models, but it’s low priority due to most of those being for personal use or in academia where many have transitioned into boxes to learn ML on.
Thanks @owenhilyard for this info and the positive note! I did look at the Steam monthly HW survey as a proxy and found ~5-7% (cant remember exactly) Pascal cards, but I suspect that’s a low estimate since Steam users are probably avid games who tend to replace cards more frequently.
Is the source for the backend also in the public section of the modular repo?
Revisiting this, this morning’s nightly includes support for NVIDIA Pascal GPUs in the GTX 10XX series and Tesla P100 vintage. That includes running basic graphs. However, there are a number of intrinsics that need to have pre-7.0 compute capability paths put in place in the Mojo standard library and kernels, so you’ll see some compilation failures on specific kernels and GPU puzzle examples until those can be updated.
The good news is that the items to be updated are all in the open-source portions of Mojo, so feel free to fix those up and send us PRs for them.
Cool! Thanks for the update, Brad! This summer I bought a new laptop which has an RTX 5090 GPU, so day-to-day I’m golden. I will try it out on my ‘classic’ laptop which has the GTX 1050 GPU. Fingers crossed!!
Today I confirmed that my laptop with the GTX 1050 GPU ran puzzles P01, P04, P12, P13 with no issues. P14 emits an error, but my fork of the Puzzles repo is from late June, so there may be other issues related to compiler changes. I’m happy that Mojo/MAX can access the GPU hardware!! Thanks @BradLarson!
Ubuntu GTX 1050 GPU
MAX 25.6.0.dev2025090805
Hats off to the Modulars! Very impressed that this has really happened, and in only a few months. Where NVIDIA and PyTorch casually drop Pascal support, Mojo picks it up. Awesome!
