writings
research
- Hardware FP4 inference for speech on Blackwell
Three speech architectures on hardware NVFP4. Weights are safe everywhere, activations fail in architecture-specific ways, and CUDA graphs decide whether the speed shows up.
notes
- Lock the clocks before you trust a number
Boost clocks drift with temperature and history. Fix the frequency or your benchmark measures the weather.
- A minimal nsys recipe for CUDA kernels
The four nsys flags that earn their keep on kernel work, and the bandwidth arithmetic to run before trusting the trace.