The problem
Traverse Research works on GPU inference and training for rendering workloads. We connect model development to execution in the renderer, including resource layout, scheduling and hardware capabilities.
The model is only one part of the frame cost. Feature generation, tensor layout, transfers and synchronisation can change the result substantially. Hardware execution paths also differ. We investigate those relationships in working systems instead of treating a standalone model benchmark as the full answer.
What we do
Breda contains neural operators, tensors, optimisers, inference support, ONNX interchange and render-graph integration. Its neural-material and radiance-cache systems exercise those components in rendering contexts.
Traverse also maintains public Rust bindings for Arm’s Neural Graphics SDK for game engines, bringing Arm’s neural super sampling and frame-rate upscaling capabilities into the Rust rendering stack.
From experiments to engineering decisions
Work is scoped around a target and an acceptance method: what the model consumes, what it produces, where it runs and how its cost is measured. Partner-specific findings remain with the partner unless approved for publication.
