frame_id: /systems/sim-scalability-benchmark
How many simulations fit on one GPU
Benchmarked how Isaac Sim performance degrades as concurrent simulation instances scale on a single cloud GPU, to inform infrastructure sizing.
Problem
Before committing to a cloud GPU sizing strategy for parallel simulation instances, the team needed real data on how many Isaac Sim instances a single GPU could run reliably, and what the actual bottleneck was — GPU, CPU, or something else.
Investigation
Ran a base simulation world, generated identical duplicates within the same environment, and incrementally increased the number of active instances on an NVIDIA L40S (48 GiB VRAM, 128 GiB RAM, 8 CPU cores), logging FPS and system resource usage at each step. An empty simulation alone dropped frame rate from 120 to 60 fps, and scaling to 8 concurrent instances dropped it to 8 fps — well under the 24 fps minimum usability threshold. System usage data pointed to CPU, not GPU, as the binding constraint, isolating roughly a 4-core-per-instance ceiling.
Outcome
Produced a data-backed instance-count ceiling and identified CPU core count, not GPU VRAM, as the scaling bottleneck for concurrent Isaac Sim instances — directly shaping subsequent infrastructure sizing decisions.