4 ms·

The benchmarks they give here [1] show generic code on par with specialized libraries that are typically written in assembler. These are vastly faster than naiv

The benchmarks they give here [1] show generic code on par with specialized libraries that are typically written in assembler. These are vastly faster than naive C implementations, so your pet peeve really does not apply here. [1] https://www.modular.com/blog/the-worlds-fastest-unified-matrix-multiplication https://www.modular.com/blog/the-worlds-fastest-unified-matr... I kind of agree with the obnoxiousness of the marketing, but it pays to actually read a little of what they do before jumping to conclusions. If only because its Chris Lattner behind it, obviously not someone who is your average "modern programmer".

TheAlchemist 3y ago
I know nothing about low-level optimizations, but I love playing with different methods and benchmarks. And did try Mojo through their playground. I found it a bit more complicated than anticipated (well, I did believe it's just like python but much faster...) but also very impressive speed. It's impossible to say for sure, since you can't run the benchmark on your machine and god knows what's behind the playground. However, I compared my naive implementation to find min of a rolling mean using python with numba, pandas, polars, numpy and mojo version. Mojo version was faster than polars and numpy, which in turn were faster than pandas and numba. That's my N=1 of course, and not a reliable benchmark. But still, I find it very impressive.
spmurrayzzz 3y ago
Lattner being behind Mojo always engendered some benefit of the doubt despite my initial gripes with the marketing claims, but his recent interview with Lex Fridman [1] further emboldened my stance that if Mojo can succeed, Lattner is one of the best-positioned engineers to see it through. [1] https://www.youtube.com/watch?v=pdJQ8iVTwj8 https://www.youtube.com/watch?v=pdJQ8iVTwj8
svantana 3y ago
That article says nothing about how they achieved those numbers. Meanwhile, in the mojo docs, the matmul results are 20-50x slower for some reason: https://docs.modular.com/mojo/notebooks/Matmul.html https://docs.modular.com/mojo/notebooks/Matmul.html
aaroninsf 3y ago
It's worse than obnoxiousness; the posture they have taken engenders skepticism and distrust. When making technical claims to a technical audience, and vs Lattner leveraging reputation, the appropriate approach is clearly not this. It's to lead with conservatively stated claims, reproducible results, and all the rest of the apparatus that is collectively understood as how adults (from which I exclude bizdev) demonstrate a better mousetrap, into an environment within which the cost of switching mousetrap technology is very great indeed at every level. Big asks deserve and require better.
- Certhas 3y ago
  Meh? They are VC funded and this is the pitch to the VCs in order to get money. Obviously this is not convincing on a technical level yet. But it's also really interesting as a design. And while the code isn't open source, I for one find it pretty fascinating to see the language being iterated on in public like this. Edit: E.g. https://github.com/modularml/mojo/blob/main/proposals/lifetimes-and-provenance.md https://github.com/modularml/mojo/blob/main/proposals/lifeti... or https://github.com/modularml/mojo/blob/main/proposals/value-ownership.md https://github.com/modularml/mojo/blob/main/proposals/value-...
  - PartiallyTyped 3y ago
    For anyone interested on why a VC would fund a PL, modular attempts to build the new shovels for the AI revolution; meaning their own platform to develop models.

TheAlchemist 3y ago

I know nothing about low-level optimizations, but I love playing with different methods and benchmarks. And did try Mojo through their playground. I found it a bit more complicated than anticipated (well, I did believe it's just like python but much faster...) but also very impressive speed. It's impossible to say for sure, since you can't run the benchmark on your machine and god knows what's behind the playground. However, I compared my naive implementation to find min of a rolling mean using python with numba, pandas, polars, numpy and mojo version. Mojo version was faster than polars and numpy, which in turn were faster than pandas and numba. That's my N=1 of course, and not a reliable benchmark. But still, I find it very impressive.

spmurrayzzz 3y ago

Lattner being behind Mojo always engendered some benefit of the doubt despite my initial gripes with the marketing claims, but his recent interview with Lex Fridman [1] further emboldened my stance that if Mojo can succeed, Lattner is one of the best-positioned engineers to see it through. [1] https://www.youtube.com/watch?v=pdJQ8iVTwj8 https://www.youtube.com/watch?v=pdJQ8iVTwj8

svantana 3y ago

That article says nothing about how they achieved those numbers. Meanwhile, in the mojo docs, the matmul results are 20-50x slower for some reason: https://docs.modular.com/mojo/notebooks/Matmul.html https://docs.modular.com/mojo/notebooks/Matmul.html

aaroninsf 3y ago

It's worse than obnoxiousness; the posture they have taken engenders skepticism and distrust. When making technical claims to a technical audience, and vs Lattner leveraging reputation, the appropriate approach is clearly not this. It's to lead with conservatively stated claims, reproducible results, and all the rest of the apparatus that is collectively understood as how adults (from which I exclude bizdev) demonstrate a better mousetrap, into an environment within which the cost of switching mousetrap technology is very great indeed at every level. Big asks deserve and require better.

Certhas 3y ago
Meh? They are VC funded and this is the pitch to the VCs in order to get money. Obviously this is not convincing on a technical level yet. But it's also really interesting as a design. And while the code isn't open source, I for one find it pretty fascinating to see the language being iterated on in public like this. Edit: E.g. https://github.com/modularml/mojo/blob/main/proposals/lifetimes-and-provenance.md https://github.com/modularml/mojo/blob/main/proposals/lifeti... or https://github.com/modularml/mojo/blob/main/proposals/value-ownership.md https://github.com/modularml/mojo/blob/main/proposals/value-...
- PartiallyTyped 3y ago
  For anyone interested on why a VC would fund a PL, modular attempts to build the new shovels for the AI revolution; meaning their own platform to develop models.