Reading about recent megakernels for LLM inference brought me back to persistent threads and megakernels I used in my own GPU work. This post traces both ideas from their origins in ray tracing to modern attention kernels and full-model LLM inference

A CUDA software rasterizer that runs the graphics pipeline entirely on the GPU. The framework manages rasterization and adaptive sampling in software, generating additional work when needed while achieving interactive frame rates.

A compact representation of large volumetric datasets that supports interactive changes to the transfer function and limited changes to the viewing perspective.

We represent large, scattered data with probabilistic summaries. Effective visualizations are created by splatting anisotropic 1D, 2D, and 3D Gaussian mixture models. This compact representation supports interactive exploration of a cosmology dataset with 2.6 billion particles.

We model uncertainty in time-dependent flows with stochastic differential equations and identify surfaces that resist or enhance diffusive transport. The method avoids expensive Monte Carlo simulation while also showing the absolute scale of uncertainty, making the resulting flow structures easier to interpret.

We render large, unstructured SPH simulations directly, without first converting the particle data into a volume. Particle sampling guided by the view and local data complexity makes ray marching faster, allowing the method to scale from interactive previews to more accurate renderings with multi-phase and single-scattering effects.

Data reduction and effective visualization using a novel GPU accelerated blue noise sampling strategy.

Understanding fluid flows with a visual analysis approach and by extracting topological features.

Efficient computation of the FTLE, a topological fluid flow feature, based on GPU acceleration.