Research writeups, tutorials, and notes from the lower levels of the stack.https://coldblime.com/en-ushttps://coldblime.com/writings/fp4-speech/https://coldblime.com/writings/fp4-speech/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.Fri, 31 Jul 2026 00:00:00 GMTnvfp4quantizationttsasrblackwellhttps://coldblime.com/writings/lock-your-clocks/https://coldblime.com/writings/lock-your-clocks/Boost clocks drift with temperature and history. Fix the frequency or your benchmark measures the weather.Wed, 29 Jul 2026 00:00:00 GMTbenchmarkingcudahttps://coldblime.com/writings/nsys-minimal/https://coldblime.com/writings/nsys-minimal/The four nsys flags that earn their keep on kernel work, and the bandwidth arithmetic to run before trusting the trace.Tue, 21 Jul 2026 00:00:00 GMTprofilingcuda