Modular Mojo claims to be over 36,000 times faster than Python for AI workloads

Modular Mojo is a new programming language designed for AI developers that is said to combine the usability of Python with the performance of C with over 36,000 times the performance of Python on a matrix multiplication workload.

Modular Mojo programming language was not in the initial plan of the company but came about when the company’s founders – who focused on building a platform to unify the world’s ML/AI infrastructure – realized that programming across the entire stack was too complicated and also ended up writing a lot of MLIR (Multi-Level Intermediate Representation) by hand.

Modular Mojo vs Python matmul

The “over 36,000 times speedup” claim comes with the matmul.py script performing a 128×128 matrix multiplication in Python with a throughput of 0.00215 GFLOP/s and another script doing 512×512 vectorized + parallelized matrix multiplication in Mojo at 79.636 GFLOP/s.

The claim looks dubious and that’s odd they used different matrix sizes, but some are ready to spend a lot of money on the company, as Monitor claims to have raised $100 million dollars. A discussion thread on Twitter/X reveals the results may not be that relevant because as Sachin Shekhar puts it:

All ML libraries use C under the hood. Python is merely an interface. Given, C is still being used, it isn’t easy to replace a programming language. You need to give something special. Nobody’s running big loops using native Python in production.

We can’t reproduce the test above just yet, because the Mojo SDK will only be released in September, but the documentation is available and includes a such about matrix multiplication. Since Mojo is a superset of Python, both matmul scripts are identical:


They just changed the function name to “matmul_untyped” in the Mojo script. Here’s the code for the Python matrix multiplication script:


The throughput is shown to be 0.0016717199881536883 GFLOP/s in the documentation.

Now we can look at the Mojo script which also imports modules specific to Mojo:


The output from the script is:


So when using the same 128 x 128 matrix, Mojo is 17.5 times faster than Python so the 36,000 times speedup claim feels fake to me as it’s only due to the larger matrix, and makes Monitor look dishonest… A 17.5 times speedup is still pretty good, but if most Python programs for AI workloads indeed rely on C libraries, it may not be relevant.

Since Mojo is a superset of Python, I suppose it won’t run on low-end hardware like microcontroller boards, but should be fine on any Linux platforms.

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

ROCK 5 ITX RK3588 mini-ITX motherboard
Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
6 Comments
oldest
newest
Arne
Arne
10 months ago

If they had something actually worthwhile, they would benchmark against some commonly used BLAS libraries like Intel MKL or OpenBLAS instead of a hand-coded, completely unoptimized Python version.

Willy
10 months ago

If the goal was to prove how horribly slow python is, everyone knows it, that’s probably even why they chose the name of a snake for the language given that it cannot run as having no legs. But in AI workloads, 95-99% of the time is still spent in underlying calculation libs. There’s definitely a bit of fat left due to the numerous python modules imported and passed through, but this is small compared to the rest. It’s likely that they wanted so make some buzz to raise funds, and if it worked, well, good for them. By the way… Read more »

TonyT
TonyT
10 months ago

Python is named after Monty Python, not the snake.

persondb
persondb
10 months ago

As you said it, it’s really extremely odd as Python really isn’t used like that all.

Python is really the interface and only does stuff that isn’t performance dependent. It is of course, a really big help, because it’s much faster to iterate over it than something like C, but still… No AI or anything really is implemented in pure Python.

There is also some other questions. Is the performance increased over CPython? Python has different implementations that have actual different performance profiles, if you want fast Python code, you will use PyPy.

Konstantin
Konstantin
10 months ago

What is the market for it? Any reasonable software enginner will do fast matrix multiplication with tensorflow or pytorch, or at least numpy/cupy. Also, benchmarking code with memory unsafe tricks against memory safe code is a questionable practice.

Khadas VIM4 SBC