Live status / Space
Space
The space study asks what the public satellite catalog can measure: the density of the upper atmosphere, read from how orbits decay; close passes and their collision chances; and how each object behaves. Most of its runs are CPU work, and they go to free queues wherever those start.
-
Upper-air density: the seed round scored 0.874
ScoredThe density model predicts how dense the upper atmosphere will be, from the sun's and the geomagnetic field's activity and from what the catalog's orbits show. Over 927 held-out days, its error a day ahead is 0.874 of the error of NRLMSIS 2.0, the standard empirical model, about 13 percent lower. A version given only the solar and geomagnetic inputs did as well (0.862), so at this lead the catalog's orbits add nothing yet, and that pass line is not met.
Figure 1. The density model's error as a share of a reference model's error on held-out days (925 to 930 days per lead). Below 1, the model is better than the reference. Persistence repeats the last measured density; it beats this model at 6 and 24 hours. The 24-hour figure of 0.874 is the main model; across five seeds it runs from 0.872 to 0.981, mean 0.902. Scored 12 October. What the catalog check means
The pass line asks whether the drivers-only model's error ratio at 24 hours is worse than the full model's by at least 0.05. It came out 0.012 better instead (0.862 against 0.874). The catalog may still help at other leads, or with a model built to use it differently; this round says it does not help at 24 hours as built.
-
AMD graphics nodes: open to every study
DoneA node with four AMD MI300A accelerators, in a free queue, installed PyTorch 2.9.1 built for ROCm 6.4 from packages we carried in, found all four devices and multiplied matrices at about 55 trillion operations a second in single precision. Every study can now use that free queue for GPU work; it sat unused for GPU work because no study had a ROCm build of PyTorch there.
The first try
The first check stopped before PyTorch ran, because it asked for NVIDIA's monitoring tool on an AMD node. The check was fixed to ask ROCm instead, and the rerun on 12 October passed.
-
Close-pass atlas and object behavior: reruns finished
Being readThe close-pass atlas (its fourth version) and the object-behavior study (its third) ran again on CPU nodes in a free queue and finished on the morning of 12 October. Their results are being read against their pass lines.
Close-pass atlas
Object behavior
-
Every orbit study, by pass line
As of 8 OctoberA pass line is the result a study must reach to count as a success. Across the orbit studies, 36 of 95 pass lines were met when they were last summed up.
Figure 2. All 95 pass lines across the orbit studies on the supercomputer, by how each read when last summed up, 8 October. The longer write-up lists each study and its lines.