> This kernel archives a bandwidth of 1056.08 GB/s which is faster than the 875.46 GB/s we archived using CUDA. I believe that to be the reason because we use the PTX api for TMA transfers in Mojo.
I can't say for sure because I couldn't find the CUDA kernel but I kind of doubt this is true. You can hit memory bandwidth on Hopper without using TMA at all, which is mostly designed for accelerating asynchronous copies and reducing memory pressure. If all you are doing is a transpose you don't need any of this to go fast (though it might simplify your indexing code…?)
I'm not an expert in this space, but is this meaningful? I'd assume that it's more common to fuse together transposition with an operation that precedes or follows it (e.g. matmul), which should be far more efficient than materializing the entire transposition in memory if it's just an intermediate value.
Matrix transpose is a canonical example of a memory bound operation and often used to showcase optimization in a particular programming language or library. See for example the cutlass matrix transpose tutorial from Jay Shah of flash attention 3 paper: https://research.colfax-intl.com/tutorial-matrix-transpose-i...
Unfortunately the issue (alluded to in the blog post you linked) is that transposes do absolutely no work but memory loads. Sure, they test that you can swizzle your accesses, but modern accelerators are all about pipelining and feeding matrix multiply units, which is considerably harder than loading from memory as fast as possible. Actually, even the Mojo post barely beats CUDA for most of its kernels, because you can hit memory bandwidth for transpose on the latest hardware using techniques from 5-10 years ago. This is definitely not true for more interesting operations.
I totally agree that the resulting kernel will be rarely useful. I just wanted to highlight that it is a commonly used educational exercise to showcase how to optimize for memory throughput. If the post showed how to fuse a transpose + rmsnorm epilogue to a gemm then the kernel would be more functional but the blog post would be much harder to follow for newcomers.
Jay Shah’s later articles contain examples that involve epilogue fusion. IMHO, understanding how to write an efficient transpose helps with following the more involved ones.
You're right that a good graph compiler will do this for you. There still may be times, like if you're interfacing with another library, where you'll need to switch a matrix between row major or column major layouts.
The next operation might need the data in column major order to read it fast. So you might have to transpose first. And these maybe be concurrent stages of a processing pipeline.
Now I'm curious, how many times do you have to fully read the matrix in GPU for the total impact of reading columns to be higher than one-off actual transpose and then sequential row reads? I know it depends on lots of things, I'm after a rough estimate.
It's quite rare. Usually problems are tiled anyway and you can amortize the cost of having data in the "wrong" layout by loading coalesced in whatever is the best layout for your data and then transposing inside your tile, which gives you access to much faster memory.
Modular (the company behind Mojo) uses it in production. I imagine that if they have any clients then those also use Mojo in production - albeit indirectly - since all the GPU kernels used by Modular are written in Mojo.
I work on Mojo. The whole compiler, runtime etc. will get open sourced, most likely within a year. It is just a matter of time and us getting all the required work done.
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%.
Users of the site only have one control available: the flag. There's no way to object only to the title but not to the post, and despite what you say that title hit the trifecta: not the original title, factually incorrect, and clickbait. So I'm not that surprised it got flagged (even if I did not flag it myself).
Email the mods at hn@ycombinator.com. There's a chance they'll remove the flag and re-up the post.
I think the OP based the title off of "This kernel archives 1437.55 GB/s compared to the 1251.76 GB/s we get in CUDA" (14.8%) and not the final kernels for whatever reason
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.
I don't think that's fair. The article promised a highly efficient kernel and seems to have delivered exactly that, which isn't "nothing". My beef is entirely with the submitted title.
I can't say for sure because I couldn't find the CUDA kernel but I kind of doubt this is true. You can hit memory bandwidth on Hopper without using TMA at all, which is mostly designed for accelerating asynchronous copies and reducing memory pressure. If all you are doing is a transpose you don't need any of this to go fast (though it might simplify your indexing code…?)
Jay Shah’s later articles contain examples that involve epilogue fusion. IMHO, understanding how to write an efficient transpose helps with following the more involved ones.
Isn't it better to simply combine the transposition with whatever next operation one wishes to do with the matrix?
>(2771.35/2775.49 - 1) * 100 = -.14916285052369131300
Flagged.
"This kernel archives 1437.55 GB/s compared to the 1251.76 GB/s we get in CUDA" (14.8%) which is still impressive
https://docs.modular.com/mojo/faq/#open-source
Are you talking about your libc equivalent or MAX?
Also, the improvement is 0.14%, not 14% making the editorialized linkbait particularly egregious.
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%.
Email the mods at hn@ycombinator.com. There's a chance they'll remove the flag and re-up the post.
> "From the moment I understood the weakness of my flesh, it disgusted me. I craved the strength and certainty of steel."
14% all the time vs 35% some of the time
edit: Closing numbers are far less impressive than those buried in the middle of the post. Confusing; bye everyone