The main result came from an end-to-end test of three ways to divide work across GPUs: a fixed split, a reactive rule and the RNN controller. The 100,000-iteration simulation modeled 435,000 cells on eight GPUs. Mean global timing imbalance—a measure of how unevenly the GPUs shared the work—was 3.5% with RNN-guided control, compared with 11.3% for static partitioning and 3.8% for the reactive rule. Total runtime was reported as 9.0% lower than the static case, while the RNN moved 7.7 times fewer spatial slices than the reactive approach.
The work is not evenly distributed
These simulations divide a three-dimensional model into spatial regions, but the amount of work assigned to a region is not treated as a question of geometric size alone. The framework combines GPU-parallel force evaluation, spatial binning—organizing elements by location for the interaction calculation—distributed spatial domain decomposition and workload-aware initial partitioning for a subcellular-element simulation. For that first division, it uses a workload score based on the number of subcellular elements and the neighboring elements within the interaction cutoff. In other words, the starting map is based on expected interaction work as well as where the regions sit in space.
A small controller with a memory
The controller itself is compact. It uses a single shared recurrent cell across interior boundaries, while keeping a hidden state for each boundary. A gated recurrent unit, or GRU, is a recurrent network that carries information from earlier steps; the reported design has one layer, a 24-unit hidden state and about 2.8 × 10^3 trainable parameters. The aim, as framed by the study, is to use that history when adjusting boundaries as the workload changes.
Training on synthetic workloads
Training happened offline in a differentiable simulator rather than from measured execution traces. The synthetic workload evolution included growth events, bursts and measurement jitter, giving the controller varied sequences to process. The reported protocol used 50 rebalancing events over 5,000 simulation iterations, with a rebalancing cadence of 100 iterations. Evaluation included unseen 100-event rollouts and an additional 400-event long-horizon rollout; scenarios ranged from two to 16 ranks. The computational tests used the epidermal SEM model on the Swan cluster, with one GPU per simulation rank, OpenCL for GPU computation and MPI for communication.
First, the engine had to be fast
Separate tests measured the interaction engine alongside the balancing comparison. At 1.6 × 10^5 elements, serial CPU interaction evaluation took roughly two hours. GPU all-pairs took 14.1 seconds, while the binned GPU version took 7.2 seconds; the reported CPU comparison was a speedup of roughly three orders of magnitude. At the largest configuration, binned GPU runtime was 255 seconds, against 2,355 seconds for GPU all-pairs, a reported 9.2-fold speedup.
Scaling has a ceiling
Those gains did not translate into ideal scaling as more GPUs were added. In fixed-size strong-scaling tests—the same problem spread across more GPUs—the larger case reached approximately 45% computation-only efficiency at eight GPUs and 31% at 16. The smaller case reached 36% at eight GPUs. The authors identify a practical constraint: at high GPU counts, a fixed-size problem can leave too little work for each GPU. Larger production simulations, including million-cell-level tissue models, remain future work.
Fewer moves, not always the lowest imbalance
On held-out rank-count scenarios, the controller’s advantage was not universal. At 16 ranks, a persistence-based forecast moved 6,481 slices and a Holt-based forecast moved 5,243, while the reactive and RNN controllers kept imbalance at or below the static method with fewer than 60 migrations. Across rank counts, the RNN had the fewest migrations among adaptive methods, but at moderate counts it was not always the method with the lowest imbalance. The reported pattern is a trade-off: reducing boundary moves did not guarantee the lowest imbalance in every tested setting.
What the test does not establish
The modeled biological setting was embryonic epidermis, shown through snapshots from Day 0 to Day 5. The authors explicitly present that sequence as an example of an evolving computational workload, not as a separate biological validation study. The document is an arXiv version 1 preprint dated 26 Aug 2026. Larger production-scale simulations, including million-cell-level tissue models, are identified as future work, so these results should be read as evidence from the tested computational setup rather than as validation of real epidermal development.
Paper data and sources
Original title: Scalable Multi-GPU Simulation of 3D Multicellular Growth with RNN-Based Workload Balancing
Authors: Matvey Moisseyev, Huijing Du, Dandan Zheng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text