Skip to content
Global sound
Creepybits

Creepybits

Quantization vs Quality Degradation

· 45 min

Show notes

What really happens when we compress AI models? In this episode, we break down the mechanics of quantization and quality degradation. We explore why FP32 is essential for training but complete overkill for inference, and unravel the paradox of why smaller GGUF files run significantly slower than FP8 and NVFP4 on modern GPUs. Finally, we put GGUF and NVFP4 to the ultimate test in video generation, wrapping up with a look at pushing 0.4 megapixel video to 1080p in real-time using NVIDIA's RTX upscaler. Is the loss of precision just a myth?

Website: zanno.se

Live, last hour: 6,913 songs on 2,569 stations in 112 countries

Discover the world through radio

More ways to discover (9)

Heard something good? Save it, or open it straight away in Spotify.

Radio 538 luisteren - Live!Popular right now