Top all-time · #9680
Pranjal Shankhdhar
Personal blog · cudaforfun.substack.comprogramminghardware
Pranjal Shankhdhar writes Pranjal’s Substack, a personal blog at cudaforfun.substack.com about programming and hardware. Since 2024, 8 different Hacker News readers have linked it 12 times, most often “Outperforming cuBLAS on H100: A Worklog”. It is still writing: the last post was on 2026-08-15, and you can follow it by feed.
My personal Substack. Click to read Pranjal’s Substack, by Pranjal Shankhdhar, a Substack publication with hundreds of subscribers.
Most-linked posts?
- Outperforming cuBLAS on H100: A Worklog6 people · 10 mentions · first linked 2024-11-30
- Outperforming cuBLAS on NVFP4: Blackwell2 people · 2 mentions · first linked 2026-08-16
What Hacker News says?
I know a trick, although it relies on the closed source cuBLAS library. The matrix multiplication in non-batched forward passes is really matrix-vector multiplication, but you can shoehorn that into matrix-matrix multiplication by setting the right dimension to 1. cuBLAS uses column major ordering to match Fortran instead of the row major ordering of C/C++, although you can swap transa and transb, m and n, a and b,…
Comment linking /p/outperforming-cublas-on-h100-a-worklog
on “Fast LLM Inference From Scratch (using CUDA)”Submitted Outperforming cuBLAS on H100: A Worklog as a story · 4 points
HN discussionthanks so much for reading!! >I think around that microarchitecture you want at least a two-deep pipeline where global memory loads to shared memory two iterations head I agree, if I were to spend more time on this, I would add another layer of pipelining like you say, and also tweak it so that each threadblock is persistent on a particular SM, and processes multiple tiles. This allows you to hide the write latency…
Comment linking /p/outperforming-cublas-on-h100-a-worklog
on “How to Write a Fast Matrix Multiplication from Scratch with Tensor Cores (2024)”Submitted Outperforming cuBLAS on H100: A Worklog as a story · 3 points
HN discussionGreat article — and several other high-quality deep dives linked at the end! Here's another one on the H100 that I found particularly useful: < https://cudaforfun.substack.com/p/outperforming-cublas-on-h1... > I agree with the author that programming GEMM on newer GPUs is a very different experience, though I'm wondering if "newer GPUs are [actually strictly] better"? It seems like there should still be some highly…
Comment linking /p/outperforming-cublas-on-h100-a-worklog
on “How to Write a Fast Matrix Multiplication from Scratch with Tensor Cores (2024)”Submitted Outperforming cuBLAS on H100: A Worklog as a story · 3 points
HN discussionSubmitted Outperforming cuBLAS on H100: A Worklog as a story · 3 points
HN discussionSubmitted Outperforming cuBLAS on H100: A Worklog as a story · 3 points
HN discussionSubmitted Outperforming cuBLAS on NVFP4: Blackwell as a story · 2 points
HN discussionSubmitted Show HN: Outperforming cuBLAS on H100 as a story · 2 points
HN discussionSubmitted Outperforming cuBLAS on NVFP4 as a story · 1 points
HN discussionSubmitted Outperforming cuBLAS on H100: A Worklog as a story · 1 points
HN discussion
