pytorch consulting the math library heuristics which say transposing the matrix will yield a 30% speedup over the non-transposed case, deciding to transpose the matrix, and immediately running out of memory
GPT-5 Pro found a counterexample to the NICD-with-erasures majority optimality (Simons list, p.25). simons.berkeley.edu/sites/default/… At p=0.4, n=5, f(x) = sign(x_1-3x_2+x_3-x_4+3x_5) gives E|f(x)|=0.43024 vs best majority 0.42904.
I encourage you to read this article, in which we describe the current situation and the directions in which, in our view, mathematics is heading. Many thanks to Ken Ono for including me in this extraordinary project. I look forward to a wide-ranging discussion and will be…
Makes sense. Mojo gives you the full power of the hardware, it doesn't "abstract" it like some other systems, so it is perfect for doing this sort of work. It provides helper libraries that you can optionally use to make some things (incl tiling etc) more declarative, and…
Hoping to see a light in one of the darkest corner of math i.e. ABC conjecture/dispute at some point in my lifetime!
Great @SimonsFdn talk by Kevin Buzzard on math's future; he notes mathematical ideas have grown so complex that traditional writing methods struggle to cope. He demonstrates how combining LLMs with proof assistants like Lean could eliminate hallucinations—the LLM proposes ideas…
AI as a multiplier has created a lot of fake talents. Take it away, they collapse. For real talents, they might become less productive!
Mathematics. The beauty of September 27, 2025. By Jerome White, @talljerome, Used with permission.
AI workloads need to run across a growing number of hardware architectures, in the datacenter & at the edge. This ‘industrialization of inference’ demands making every FLOP count through increased resilience & lower costs. USIT believes @Modular is best positioned to become…
We are very thankful to all our new and exiting investors who backed us. This round was lead by Thomas Tull's @usitfund, with @dfjgrowth & all existing investors: @gvteam @generalcatalyst @greylockvc investing. Read more ⬇️ modular.com/blog/modular-r…
Let's gooooooo Modular 🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀
We raised $250M to accelerate building AI's unified compute layer! 🔥 We’re now powering trillions of tokens, making AI workloads 4x faster 🚀 and 2.5x cheaper ⬇️ for our customers, and welcomed 10K’s of new developers 👩🏼💻. We're excited for the future!
We beat Nvidia’s cuBLAS kernels on B200s in 170 LOC. Using zero CUDA. Just pure Mojo. Here’s exactly how we went from 1% to 106% of Nvidia benchmark perf from scratch (with code) 👇🧵
We're going to devour llama.cpp, mlx, etc. one step at a time!
We're thrilled to announce Modular Platform 25.6 🚀 Our biggest step yet toward a unified compute layer for AI. • Peak performance on NVIDIA Blackwell & AMD MI355X • Mojo on Apple, AMD & NVIDIA consumer GPUs modular.com/blog/modular-2…
We know that one of the biggest barriers to programming GPUs is access to hardware: "Code you’ve written for NVIDIA or AMD GPUs should now mostly just work on an Apple🍎 Silicon GPU, assuming no device-specific features were being used." Preview here:👇 forum.modular.com/t/apple-silico…
Also the code for part 1-4 are here github.com/modular/modula… will add to the blog posts too.
Part 4 of "Matrix Multiplication on Blackwell" is here! It continues our epic journey of describing how Modular implemented the fastest B200 matmul in the industry, revealing the techniques to achieve 1772 TFLOPs, exceeding that of the current SOTA. modular.com/blog/matrix-mu…
Part 4 of "Matrix Multiplication on Blackwell" is here! It continues our epic journey of describing how Modular implemented the fastest B200 matmul in the industry, revealing the techniques to achieve 1772 TFLOPs, exceeding that of the current SOTA. modular.com/blog/matrix-mu…
torch.compile is PT's achilles heel!
Compiling large #PyTorch models at Meta could take an hour+. Engineers cut PT2 compile time by 80% with parallel Triton compilation, dynamic shape marking, autotuning config pruning, and cache improvements now integrated into the stack. 🔗 hubs.la/Q03J-6P20
Rich guy open source passion projects like Ghostty and Omarchy are the internet equivalent of building a library or school
Triton is nice if you want to get something onto a GPU but don't need full performance/TCO. However, if you want peak perf or other HW, then Mojo🔥 could be a better fit. I'm glad OpenAI folk are acknowledging this publicly, but I wrote about it here: modular.com/blog/democrati…
TIL, RIP Triton, killed by inability to have good Blackwell performance
United States 趋势
- 1. Bears 89.5K posts
- 2. Jake Moody 13.8K posts
- 3. Snell 24.6K posts
- 4. Caleb 48.9K posts
- 5. Bills 142K posts
- 6. Falcons 51.5K posts
- 7. Happy Birthday Charlie Kirk 5,308 posts
- 8. phil 178K posts
- 9. Josh Allen 26.7K posts
- 10. #BearDown 2,389 posts
- 11. Jayden 22.9K posts
- 12. Joji 30.6K posts
- 13. Swift 289K posts
- 14. Ben Johnson 4,437 posts
- 15. Troy Aikman 6,590 posts
- 16. #Dodgers 15.4K posts
- 17. Turang 4,348 posts
- 18. Roki 6,033 posts
- 19. Bijan 33.1K posts
- 20. #BeckyEntertainment 103K posts
Something went wrong.
Something went wrong.