Independent worlds never collide with each other. Fused ensemble mode shares one solver graph and GPU storage; other modes retain separate world graphs. Use the same body count and workload when comparing batched and separate submission. Use one world to measure single-world latency.

Each world uses local float32 coordinates and a double-precision global origin. This example places world origins one billion units apart. Resident memory includes buffers, GPU reset snapshots and graph scratch, excluding pipelines and driver overhead.

Timing covers encoding, submission and GPU completion. Compilation, diagnostics, final-state drawing and recycling are reported separately or excluded. A successful completion flag does not establish physical accuracy. Ensembles share one rollback boundary and contact budget, retain one template reset image, and currently exclude joints, CCD and authored triggers. At most 65,535 bodies and colliders fit in one ensemble. Inactive worlds retain state. Throughput counts only active world/body steps; the actual active count is rounded up. GPU scheduling remembers sleeping instances. Contact lists, rollback payloads and bucket clearing follow active work. Scheduling and status still scan instance records. Mostly sleeping tests quiescent residency; Wake storm wakes enabled instances every eight frames. GPU stage timings come from a separate instrumented step after the run, outside throughput measurements. Optional events and publication can still scan capacity. Recycling applies only to separate-world modes.