timmyd

Updated -> https://mojolang.org/.

"Why Mojo?" is now smack bang on the homepage!

You'd likely enjoy: https://mojolang.org/nightly/docs/vision/ for why we created it, and there's https://mojolang.org/nightly/docs/faq/ as well :)

Feedback welcome as always and thanks for checking it out! Encourage you to watch ModCon next week for some exciting announcements :)

David Robertson took a quantization challenge designed for CUDA experts, and solved it in Mojo with AI assistance, and ended up 1.07x to 1.84x faster than the state-of-the-art C++/CUDA implementation.

Co-founder here. There isn't any signup - that was 2+ years ago and we've been iterating a lot with the community and listening to feedback - which has been wonderful. Go freely and install with Pip, UV, Pixi etc -> https://docs.modular.com/mojo/manual/install

You might also enjoy this series: https://www.modular.com/democratizing-ai-compute which goes into a lot of the details.

thanks jsnell - i did they and they appreciated the comment above, and unflagged it. i appreciate it!

FWIW I didnt take the blog as a dunk on CUDA, just as an impressive outcome from the blog writer in Mojo. It's awesome to see this on Hopper - if it makes it go faster thats awesome.

[op here] To be clear: Yes, there are 3 kernels - you can see those in the linked github at the end of the article if you clicked that. These are:

transpose_naive - Basic implementation with TMA transfers

transpose_swizzle - Adds swizzling optimization for better memory access patterns

transpose_swizzle_batched - Adds thread coarsening (batch processing) on top of swizzling

Performance comparison with CUDA: The Mojo implementations achieve bandwidths of:

transpose_naive: 1056.08 GB/s (32.0025% of max)

transpose_swizzle: 1437.55 GB/s (43.5622% of max)

transpose_swizzle_batched: 2775.49 GB/s (84.1056% of max)

via the GitHub - simveit/efficient_transpose_mojo

Comparing to the CUDA implementations mentioned in the article:

Naive kernel: Mojo achieves 1056.08 GB/s vs CUDA's 875.46 GB/s

Swizzle kernel: Mojo achieves 1437.55 GB/s vs CUDA's 1251.76 GB/s

Batched swizzle kernel: Mojo achieves 2775.49 GB/s vs CUDA's 2771.35 GB/s

So there is highly efficient matrix transpose in Mojo

All three Mojo kernels outperform their CUDA counterparts, with the naive and swizzle kernels showing significant improvements (20.6% and 14.8% faster respectively), while the final optimized kernel achieves essentially identical performance (slightly better by 4.14 GB/s).

The "flag" here seemed innapropriate given that its true this implementation is indeed faster, and certainly the final iteration could be improved on further. It wasn't wrong to say 14% or even 20%.

Updated the title to the original. I did base the numbers on

"This kernel archives 1437.55 GB/s compared to the 1251.76 GB/s we get in CUDA" (14.8%) which is still impressive

I do find it amusing that all comments are focused on the US. There are hundreds of countries around the world were TV stations are black boxes and presidential / prime ministerial hopefuls have no platform to express themselves. Having access to a platform that can promote and answer content as users are searching for answers on them (at least how it reads on face value) seems incredibly important. I do however agree that transparency around who gets access to the platform is critical for user trust.

Well consider the number of 3rd party services that would shut or require a significant amount of development to get back online as a result of google shutting down:

https://developers.google.com/apis-explorer/

There's a lot of services that depend on Google and changing all that code overnight is unlikely. Apple doesn't have that kind of 3rd party reach.

appreciate any feedback, insight etc you guys can give us! we're working hard to spread this across the bay area