The biggest gap appeared as clusters grew
DISHTINY, one of the study's digital-evolution workloads, retained 92% scaling efficiency as its process count rose from 1 to 64 under best-effort execution, compared with 47% for synchronous execution. Beyond 16 processes, the synchronous result degraded by 34%, while the best-effort result remained stable. Here, best-effort execution meant the processes did not all have to wait in lockstep. The reported Mann-Whitney test gave a p-value below 0.001. These figures describe the tested workload and setup.
The graph-coloring workload showed a similar gap. At 64 processes, the reported speedups were 2.1-fold for DISHTINY and 12.5-fold for graph coloring. Graph-coloring solution error was 73% lower in the comparison. All three comparisons were reported as statistically significant, with p-values below 0.001.
On the CPU cluster, each process handled 3,600 DISHTINY cells or 2,048 graph-coloring nodes. There were 10 independent executions per SLURM allocation, or cluster job allocation. The study therefore compared repeated computational trials rather than a single pooled sample.
Typical service held up, while the extremes widened
Speed was only part of the test. The cluster analysis measured quality of service, or QoS, across four runtime dimensions: straggling, latency, attrition and bunching. Higher values meant worse service.
When graph-coloring allocations increased from 64 to 256 processes, every reported QoS metric developed more extreme tail values, and all comparisons had p-values below 0.001. Yet median QoS showed no detectable degradation, and straggling fell 4% in the main comparison. Under extreme communication intensity, median straggling instead worsened by 8%.
That distinction matters because a typical run can look steady while the worst cases become more severe. The findings do not show that stable median QoS rules out rare, severe runtime disruptions.
Averages did not explain every slowdown
One test focused on a piece of hardware identified as lac-417. Allocations that included it produced extreme outliers, but median QoS remained stable across all metrics. Median straggling was 0.5% lower when lac-417 was included, with a reported p-value of 0.04. That association should not be read as evidence that the anomalous hardware improved the runs.
Another diagnostic compared processes placed across nodes with processes placed within the same node. Incoming and outgoing traffic were anti-correlated in both configurations, with a generalized linear model reporting a p-value below 0.001. The across-node imbalance coincided with a 12% host-node execution-speed disparity. After that disparity was taken into account, no detectable traffic imbalance remained, with a p-value of 0.15. For imbalance within a node, the paper points to NUMA, a hardware layout effect, as a likely explanation.
At wafer scale, the history was necessarily approximate
The study then moved to a third-generation wafer-scale engine, or WSE, arranged as a 755 by 1,170 lattice of 883,350 processing elements. The platform was described as having 44 GB of aggregate memory and 48 kB per element. Its two case studies involved tracking the history of dummy agents and following the evolution of a mutator allele.
For the history project, the researchers used single-bit barcode markers in a fixed 64-bit region of each genome. Once that space filled, new markers overwrote older ones. Known injection times then supported a trie-based reconstruction of the phylogeny, or branching ancestry record. This was an estimate built from markers, not a complete record of every generation.
In one WSE run, the system evaluated 226.1 million agents for 5 million generations at about 8,750 generations per second. Asynchronous device-to-host transfers sampled one genome per processing element about 15 times per second and produced 9,000 snapshots. Those figures describe a sampled record, not a complete evolutionary history.
A transfer fault hit records while the simulation kept running
During a localized device-to-host transfer fault, about 5% of copied 32-bit words in the affected region had bit flips, but only 0.03% of extracted genomes were corrupted. Parity checks indicated that the on-device simulation was unaffected, and the work continued uninterrupted for two months.
That result indicates resilience for this localized event and workload, not universal fault tolerance. The reported evidence does not establish how the method would perform under other hardware anomalies or across other HPC workloads.
The monitoring also exposed possible bias
The WSE work also looked for bias in population dynamics. Under neutral conditions, reconstructed lineages were expected to behave like spatial random walks. Instead, parallel lineages followed similar trajectories, and many moved from the bottom of the chip to the top. The pattern suggested directional migration imbalance.
In the mutator-allele project, fixed-size single-bit buffering supported recordings of unknown duration by overwriting and downsampling data over time while keeping memory use bounded. Mutator prevalence transiently reached up to 99.9%, even when fixation was unreliable.
A narrower claim about future computing
Taken together, the tests support a scoped argument for best-effort computing in digital evolution. The reported gains came with a need to watch both QoS tails and domain-specific signs of computational bias. The authors argue that digital evolution has a unique opportunity to contribute to post-deterministic computing and to help adapt evolutionary research to emerging accelerator architectures.
The evidence covers the reported DISHTINY and graph-coloring CPU-cluster workloads and the WSE evolutionary simulations. It does not establish general superiority across all HPC workloads, accelerator platforms or hardware faults, and the WSE sampling strategy does not provide complete evolutionary histories. Confidence intervals were not reported. The manuscript is an arXiv preprint, version 2, dated 3 Sep 2026.
Paper data and sources
Original title: Trust, but Verify: Rigorously Profiling Best-Effort High-Performance Computing for Digital Evolution
Authors: Matthew Andres Moreno, Santiago Rodriguez Papa, Charles Ofria et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text