Analyzing Deepseek's "undefined" NVIDIA PTX optimizations (with benchmarks!)
Published February 27, 2025
Views 131K | Likes 7.37K | Comments 576
At a glance
Views
131,480
Likes
7,379
Comments
576
Performance vs. channel
Outlier score
Engagement rate
More features coming
We're working on new analytics and tools. Stay tuned.
About
Two days ago, Deepseek surprised everyone with an "undefined-behavior" PTX optimization speeding up particular ML workloads on a Hopper NVIDIA GPU Kernel. Let's reverse engineer the hack, implement it ourselves, and benchmark the speedup on an H100. -- Link to my test code: https://github.com/LaurieWired/BenchmarkCu…
- Published
- February 27, 2025
- Made for kids
- No
More features coming
We're working on new analytics and tools. Stay tuned.