GPU determinism idea: Field‑Oriented Control for GPUs

Field-oriented control is the feedback system that keeps an electric motor smooth by measuring it constantly and correcting it in real time. Engineer Ben Coupland said on July 8, 2026 that the same closed-loop idea could be applied to graphics processing units, or GPUs. (mathworks.com) (github.com) A GPU usually chases maximum throughput, not a fixed schedule, and that makes repeatable timing hard when thousands of threads compete for shared resources. NVIDIA said on March 5, 2026 that deterministic floating-point results on GPUs often require slower reduction methods because parallel math and scheduling do not naturally produce identical outcomes every run. (developer.nvidia.com) Coupland’s prototype, published as the GitHub project GPUtronic, treats the GPU as an “Instruction Pressure Engine” and adds a governor that adjusts work in real time. The repository describes a proportional-integral-derivative controller, dynamic block scaling, and “thermal knock retard,” borrowing language from engine control units and motor drives. (github.com 1) (github.com 2) Field-oriented control works by splitting one hard control problem into simpler parts, then closing the loop with constant measurement and correction. MathWorks describes the method as a stack of current, speed, and protection loops designed to hold torque and speed steady across changing conditions. (mathworks.com) The GPU version is aimed at a different target: deterministic multithreading for workloads that must finish predictably, not just quickly. Coupland linked that idea to aerospace and onboard perception, where sensor-heavy software has to balance large parallel workloads with timing and safety constraints. (github.com) That pitch lands in a field where determinism is still partial and workload-specific. NVIDIA’s CUDA Core Compute Libraries now let developers choose among “not guaranteed,” “run to run,” and “GPU to GPU” determinism for some reductions, while NVIDIA’s reproducibility repository documents similar limits in machine learning frameworks. (developer.nvidia.com) (github.com) Other NVIDIA software makes the tradeoff explicit. PhysX 5.1 documentation says cross-platform determinism is not supported and that newer GPU-only features such as cloth and soft bodies are not deterministic. (nvidia-omniverse.github.io) Academic work is also trying to make shared GPUs behave more predictably. A March 2026 paper on a system called DetShare said it sought “absolute semantic determinism” while preserving utilization and isolation through a new scheduling layer above physical GPU resources. (arxiv.org) Coupland’s proposal is still a proof of concept, not a vendor feature or published standard. But it puts a concrete control-system frame on a long-running GPU problem: if a processor can be steered like a motor, engineers may be able to trade a little peak speed for timing they can count on. (github.com) (developer.nvidia.com)