THE CRUNCH

HPE's Spaceborne Computer 4 will ride to the Moon aboard Astrolab's FLIP rover, which is set to be deployed near the lunar south pole by Voyager Technologies' Griffin-1 lander on a mission intended to launch in late 2026 as part of NASA's Commercial Lunar Payload Services program. Assuming a safe landing, the module will spend a year running AI and high performance computing experiments, including tuned LLM tests with a retrieval augmented generation system for searching onboard documents, machine learning against camera and LiDAR data, and HPL and HPCG benchmarks. If those benchmarks run as planned, HPE and Astrolab would hold the first Top500 submission made from a celestial body.

The system itself is small: roughly 5x10 inches and about an inch thick, fanless, with a low-power CPU, a GPU and a pair of mirrored Kioxia SSDs totalling 1 TB. The actual compute silicon remains undisclosed, reportedly because HPE's compute partner does not want its name tied to hardware on a high-risk mission. More details are expected once the system powers up on the surface.

What is known is that SBC4 is largely assembled from off-the-shelf components, including standard Kioxia BG6 M.2 drives, with protection coming mainly from software rather than custom space-grade silicon. HPE's chief scientist of space tech, Mark Fernandez, said the main line of defence for storage is a software RAID 1 mirror, and that nine years of Space Station experience convinced Kioxia nothing special was needed for space. Early Spaceborne Computers did suffer high SSD failure rates, but Fernandez said the same drives failed terrestrially too, with bad capacitors rather than cosmic rays to blame, and that he has seen zero SSD failures since 2019.

Radiation and thermal extremes are still real constraints: the Moon lacks the atmosphere and magnetic field that shield terrestrial compute, there is no air for a heatsink so the rover uses a radiator, and launch and landing bring severe vibration. HPE's answer is what Fernandez calls hardware hardening in software, using spare CPU cycles to monitor hardware metrics against a state table and degrade gracefully, from slower speeds down to a controlled power-down rather than destruction. A workload like HPL may even help keep the rover from icing over during the roughly 14-day lunar night cycles it will repeatedly endure.

Beyond HPE's own experiments, NASA and Deloitte are joining the mission with additional AI workloads. NASA will run its FLUID experiment coordinating three AI models trained on data from Earth, the Moon and the Space Station, while Deloitte's Silent Shield system will provide near real-time predictive cyber-analytics and anomaly detection. HPE will also support Astrolab's mission control in Hawthorne, California, using its own private cloud stack.

WHAT HAPPENS NEXT

More details about the undisclosed compute silicon are expected once SBC4 powers up on the lunar surface, and the mission is intended to launch in late 2026.