An analysis article on Lobsters AI currently circulating in the AI underground community dissects ThunderKittens — and the comment section is starting to buzz. This is not news most people have caught onto yet, but among those who actually write CUDA kernels for a living, this is a name that is appearing more and more often.

So what's the deal? ThunderKittens is a DSL (domain-specific language) that lives within CUDA, created by Stanford's Hazy Research Lab. The idea is to provide a high-level abstraction layer that allows you to program the GPU hierarchy — warp groups, tiles, shared memory — without losing control over what is actually happening in the machine. It's a kind of middle ground between writing raw CUDA (painful, but fast) and using Triton (simpler, but with performance ceilings).

14x faster than Triton on linear attention is not a fine-tuning — it's an architectural leap.

The numbers cited from Hazy Research are brutal if they hold: FlashAttention-forward on H100 is 30% faster than FA2. Mamba-2 implementations are "several times faster" than the Triton version. For linear attention models like Based and LoLCATS Hedgehog, we're talking 14x and 6.5x speedup. ThunderKittens 2.0, which came out in February this year, claims to beat cuBLAS on B200s for BF16 and the new MXFP8/NVFP4 formats.

It's worth noting: these are figures from the lab that created the tool itself, not from an independent benchmark study. The community source here is primarily Stanford's own publications and blog posts — and there is currently no large, neutral comparative study that pits ThunderKittens, Triton, and TVM against each other on equal terms. Take the numbers seriously, but hold back a bit until replication studies emerge.

What makes this particularly interesting is not just the performance, but the positioning. Triton (OpenAI/Meta) has become the de facto standard for people who want to avoid raw CUDA, but ThunderKittens points to a real performance ceiling in Triton — especially on Hopper and Blackwell architectures where WGMMA instructions and TMA data flow are critical. ThunderKittens is built precisely for these.

If this scales and community adoption picks up, we could see a shift in how the most performance-critical AI kernels are written — especially in research environments working with new attention mechanisms and state space models. That's the space ThunderKittens clearly aims for.

Worth keeping an eye on. This is still an early signal from community sources, but the buzz is real.