Democratizing AI Compute: Q&A (Community Meeting)

Community Showcase
1

During our upcoming community meeting on March 10th, @clattner will be answering questions from the community on our recent Democratizing AI Compute series!

Please share any questions you’d like Chris to answer in this thread, or submit them here: https://forms.gle/CVwmpinDQ4yh5eNA9

DCP5
DCP51800×1024 375 KB

Part 1: DeepSeek’s Impact on AI
Part 2: What exactly is “CUDA”?
Part 3: How did CUDA Succeed?
Part 4: CUDA is the incumbent, but is it any good?
Part 5: What about CUDA C++ alternatives?

1 Like
2

One of the strategies that was outlined for CUDA’s success was that it ran and worked on developer grade hardware so developers could learn how to use it.

Currently only very expensive GPUs are supported for MAX, is Modular looking to follow a similar path to CUDA and support GPUs that people have access to sooner rather than later to speed adoption? Or will premium GPUs be the main focus? Why not start with developer grade hardware and then work up?

What lessons from the Objective C->Swift transition can be applied to CUDA->Mojo/MAX?

Not CUDA specific, but have there been any emergent features in Mojo that you particularly like?

1 Like
3

A couple of questions:

  1. Julia’s ecosystem has offered a mature GPU programming model for NVIDIA and AMD for a few years. Other than “Julia is not Python”, what’s your take on the technical reasons behind Julia not reaching the level of adoption in the AI community to be considered a viable alternative to CUDA?
  2. Many of the AI ASIC vendors like Cerebras, Samba Nova Systems, AWS Trainium, Groq, etc… have built out their own AI graph compilers and low-level kernel programming interfaces. Essentially, they have taken a page out of NVIDIA’s playbook and are building their own “CUDA” specifically for their hardware. What’s your take on how Modular’s vision fits within this growing fragmentation of the accelerated computing software stack?
1 Like
4

Question: Can you characterize in a few dot points how exactly MAX (as a project) differs from all of the previous projects that aimed to offer a hardware-independent compute platform? I’ve heard several detailed answers at this point (that includes your blog series), but not a succinct answer that I can easily share with others. :slight_smile:

In fact, it would be cool to have those dot points listed on the MAX page of the Modular website. I expect programmers would find such a comparison useful.

5

Question:

One of the factors that made CUDA so popular were the libraries provided for GPU compute, such as: cuFFT, cuRAND, and cuBLAS.

Now that we’re able to write our own GPU kernels in Mojo: which is amazing!!!

Would there be plans in the future for exporting Mojo libraries with the C ABI?

Do you think MAX/Mojo have all the ingredients required to provide a great experience for library writers?

I would love to hear any thoughts you have about this. Thank you!