Felix

All evals of ML models suck - but some are useful 🙃

Adding multi-level performance models to diagrams. This will allow performance models of FlashAttention / matmul / distributed MoEs to be dynamically calculated. Colors indicate execution at different levels, and the hexagons indicate a partitioned axis.

I solved every single problem in the CUDA mode book. A quick thread summarizing this experience and what I learned 1/x

hacker news doing hacker news things 😄

This matmul visualization is so cool it got me banned last time I posted it

open source speech synthesis model trained on two 4090 GPUs!

Qwen3 0.6b is a shockingly good draft model a lot of the time (96.6% acceptance rate on the 4b model for this particular task!)

if no one else is showing that RL isn't just eliciting latent behavior already learned in pretraining, but is actually a new scaling paradigm, nvidia has to do it themselves

though the memory bandwidth look looks pretty meh... feels like Intel is trying to make cards with @tenstorrent performance characteristics, though much cheaper(?) xD

Intel a 24GB GPU for ~500 USD, and there will be a single card version with GPU dies and 48GB combined memory

if you want people to see your post on 🦋 site then you gotta post about it on X 😂