RADIATION / SI-P1 PATHFINDER AND SI NODE
What space does
to a computer.
Above the atmosphere, protons trapped in Earth's magnetic field, cosmic rays from the galaxy and particles from the Sun pass through every spacecraft. One particle can flip a bit, stop a processor or change a result without a trace, and over the years the dose wears the chips down. Here a host satellite carrying a payload like SI-P1 flies two weeks of a real orbit through NASA's radiation-belt models and measured cosmic rays, or through the October 1989 solar storm as the GOES-7 satellite measured it. The scenarios compare the pathfinder with our SI Node, through quiet years and the storm, and show what half the shielding does; every parameter stays open for engineers.
01 / WHERE THE PARTICLES ARE
One region takes nearly every hit.
THE ANOMALY6.6% of the time, 96% of the hits.
Earth's magnetic field is offset from the planet's centre, so over the South Atlantic the inner radiation belt comes down to a few hundred kilometres. A host at 525 km in a 10:30 sun-synchronous orbit crosses it on about 11 passes a day and spends 6.6% of its time there, yet 96% of the protons that reach its electronics arrive there. NASA's AP-8 model, built on data from the 1960s and 1970s, puts the anomaly where it was then; we move it 16° west, to where the Lucky 7 CubeSat measured it in 2021, and on by 0.29° a year to 2028.
COSMIC RAYSFew, but from everywhere.
Outside the anomaly the protons come from the galaxy, from every direction the Earth does not block. Earth's field turns the slower ones away near the equator and lets them in over the poles. Behind 3 mm of aluminium they cause 4% of the hits on the 10:30 orbit. Their spectrum is the one PAMELA measured near solar minimum and AMS-02 near solar maximum; when the Sun is active, its magnetic field keeps more of them out.
SOLAR STORMSRare, fierce and polar.
On 19 to 27 October 1989, GOES-7 measured up to 42,000 protons above 10 MeV per cm² per second per steradian, one of the largest events on record, and the magnetic storm that came with it (Dst fell to −268 nT) let them reach lower latitudes. Replayed on the 10:30 orbit behind 3 mm, the nine days bring 2.3 restarts, almost as many as a quiet year, and 379 rad of dose, most of a quiet year's. A 53° orbit stays inside the field's protection most of the time and gets 30 rad.
02 / WHAT IT DOES TO THE CHIPS
Restarts, silent errors, dose.
SI-P1 · RESTARTSAbout 2.7 a year.
We found no published proton test of the Thor family yet, so the simulator uses the nearest published module: Jetson Orin NX, tested with 480 MeV protons at TRIUMF. Adding up its parts, it stopped and needed a reset about once per 3.2×10⁸ protons per cm². Behind 3 mm on the 10:30 orbit that is 2.7 restarts a year, 96% of them in the anomaly. Recovering in two minutes costs 5 minutes of service a year.
SI NODE · SILENT ERRORSOne every 4 days, unless checked.
Data-centre accelerators are not built for space. In Google's proton tests of its Trillium TPU running transformer models, a wrong result appeared without any error about once per 17 rad. An 8-chip node at 650 km in a dawn–dusk orbit, behind 10 mm, would see one every 4 days: 11 per chip a year, against 9 by Google's own numbers for its orbit. The Orin module showed no wrong result in its test, which bounds SI-P1's rate below 0.13 a year.
TOTAL DOSEDecades for SI-P1, years for a node.
Behind thin shields the dose comes mostly from trapped electrons, which a few millimetres stop; behind thick ones, from protons. The Orin module would take in 528 rad a year behind 3 mm on the 10:30 orbit, and tests saw its first failures at 37 krad: about 70 years. The accelerators' memory showed irregularities after 2 krad, which a node behind 10 mm reaches in about 9 years at solar minimum.
03 / WHAT IT MEANS FOR THE SERVICE
Numbers to design to, then to measure.
AVAILABILITYRestarts cost minutes a year.
At 2.7 restarts a year, a two-minute recovery leaves SI-P1 available 99.999% of the time, and even fifteen minutes leaves 99.992%. What matters is that a restart loses no finished work and needs no one on the ground. In a storm like October 1989 they come faster: up to one every 4 hours at the peak.
CORRECTNESSSilent errors must be caught.
A result that is wrong without any sign is worse than a restart. The standard answer is to check: run every job twice and compare, for half the compute, as the simulator lets you try, or check results more cheaply where the workload allows. How SI Runtime handles faults is not published on this website; its reliability is a target that flight and beam tests have to show.
WHAT WE WILL MEASUREReplace references with tests.
Published Orin tests differ: a 2025 summary by ESA and the Barcelona Supercomputing Center gives a functional-interrupt cross section 45 times smaller, about 7×10⁻¹¹ cm² at 200 MeV, which would mean one restart every 17 years. The simulator takes the cautious figure. Our plan includes radiation tests of the SI-P1 engineering model in months 6 to 8, before the flight is confirmed.
04 / METHOD AND LIMITS
Standard models, measured data, stated limits.
Trapped protons and electrons come from NASA's AP-8 and AE-8 models for solar minimum and maximum, with ESA's interpolation at low altitude, evaluated in the magnetic field each model was built with through the IRBEM library and stored on 2° and 4° grids at five altitudes from 450 to 650 km. The anomaly is moved to the centroid of protons above 30 MeV that the Lucky 7 CubeSat measured at 540 km on 2 January 2021 (26.6° S, 49.1° W), then west at the 0.29° a year that RXTE measured, to 20 March 2028. Cosmic-ray protons follow the interstellar spectrum of Bisschoff and Potgieter with the force-field approximation, fitted to PAMELA (2006 to 2008) and AMS-02 (2011 to 2013) to 4.9% and 2.7% rms; the Earth and its atmosphere block part of the sky, and the field turns away protons below the cutoff rigidity computed from IGRF for 2028. The storm replays GOES-7's integral proton channels and Kyoto's Dst, lowering the cutoff latitude by 0.04° per nT of Dst, the shift SAMPEX saw in the storm of July 2000. Protons are carried through the aluminium with NIST's PSTAR ranges, and dose comes from SHIELDOSE-2 at the centre of aluminium spheres.
The host's orbit is the same two-week propagation as on our request page. Event rates are each test's cross section times the flux of protons above the chosen energy behind the shielding; the globe shows as many events as expected over the mission, placed along the orbit in proportion to the rate. An independent calculation calls the models directly: the stored maps match it to 2.2% at the median and 17% at the 90th percentile, orbit averages to 5%, time in the anomaly to 8% and the storm to 0.7%. Behind 10 mm at 650 km in a dawn–dusk orbit, the model gives 216 rad a year at solar minimum and 163 at maximum, against the 150 Google estimated for that orbit. Automated tests check these limits on every build.
Not modelled: the uncertainty of the belt models themselves, often put at a factor of two (PAMELA measured fewer trapped protons at low altitude than AP-8 predicts); the direction particles come from, which matters for a spacecraft that keeps one face to the Earth; heavy ions, which upset chips more readily than protons; secondary particles made in the shielding; the exact shape of the shielding, which here is a sphere of aluminium; and how a cross section changes with proton energy, taken here as a step at the chosen threshold. Dose limits are those at which tests saw the first problems, under the tests' conditions. A communication tool, not a radiation analysis for a mission.
Sources: IRBEM library with AP-8 and AE-8 (Sawyer and Vette, 1976; Vette, 1991) and ESA's interpolation (Daly et al., 1996); Kovář and Sommer, CubeSat observation of the radiation field of the South Atlantic Anomaly, Remote Sensing 13, 1274 (2021); Fürst et al., RXTE (2009); Adriani et al., PAMELA trapped protons (2015); Bisschoff and Potgieter, Astrophysics and Space Science 361 (2016); proton spectra of PAMELA (Adriani et al., Science, 2011) and AMS-02 (Aguilar et al., Physical Review Letters, 2015) from the Cosmic-Ray DataBase; GOES-7 energetic particle data, NOAA NCEI; Kyoto Dst index; SAMPEX cutoff suppression, LASP; NIST PSTAR; SHIELDOSE-2 (Seltzer, 1994); Rodriguez-Ferrandez et al., IEEE IOLTS (2024), Jetson Orin NX; Rodriguez-Ferrandez et al., ACCEDE (2025), ESA and Barcelona Supercomputing Center; Google, Towards a future space-based, highly scalable AI infrastructure system design (2025). No chip maker, test facility or agency is represented as a partner.
