DataNVIDIA turns synthetic training data into a specification to inspect
NeMo Data Designer’s recent research and software updates move attention from generating examples to controlling what those examples teach.
NVIDIA
Model-building research and analysis covering NVIDIA.
Data
TrainingNVIDIA’s four-bit training makes precision a budget decision
Nemotron 3 Ultra shows why cheaper pretraining depends on knowing which calculations cannot afford to be cheap.
TrainingFlashAttention-4: Faster GPUs need smarter software to speed up AI training
Expensive GPUs can still spend time waiting on slower operations. Research from Together AI and five other organizations shows how better scheduling can turn more of that hardware into useful work.
InferenceSGLang’s smaller cache could make long-context inference cheaper—but accuracy sets the limit
The latest cache-compression results expose two distinct gains: faster generation and room for more live sessions.