Transformations in computer architecture enable advances in exploration autonomy
By Rick Kwan|December 2019
The Computer Systems Technical Committee works on advancing the application of computing to aerospace programs.
The architecture of terrestrial computing is in a metamorphosis that could eventually prove useful in planetary exploration. At its core is the rapid adoption of data science — not only by scientists and engineers but even by retailers sitting on stockpiles of business data. And the underlying statistics and linear algebra increasingly employ shorter precision arithmetic and logic for greater speed.
The fastest supercomputer on the Top500 list in 2019 was the U.S. Department of Energy’s Summit at Oak Ridge National Laboratory in Tennessee. Introduced in June 2018, Summit is built on a combination of IBM POWER9 central processing units and Nvidia Volta graphics processing units and is rated on High Performance Linpack at 148.6 petaflops. Beyond physics simulation, the Volta GPUs are positioned for machine learning and artificial intelligence research through the use of small integers.
The vectors and matrices of data science and machine learning aren’t necessarily single- and double-precision floating points. Much of it can be done with half-precision 16-bit floats and small 8-bit and 4-bit integers. For a given area of silicon and power budget, smaller precision allows for greater parallelism, hence increased speed.
This highly parallel reduced precision is employed in several low-end neural processing unit or tensor processing unit products released in the past 12 months. The Nvidia Jetson AGX Xavier 8G was available in August, and the Google Edge tensor processing unit was released in March. Some are targeted at mobile platforms such as drones, rovers and autonomous cars. Tesla Motors began shipping its own “Full Self-Driving” chip in March.
The neural processing and tensor processing units are an evolution of graphics processing units, which a decade ago morphed into physics engines and provided a radical boost for modeling and simulation. Neural processing units and tensor processing unit chips will eventually lead to more advanced autonomy in planetary explorers.
Summit is designed to facilitate both physics simulation and artificial intelligence research. In June, Oak Ridge reported that a Monte Carlo simulation of a modular nuclear reactor ran 30 to 40 times faster on Summit than on the Oak Ridge’s Titan supercomputer introduced in 2012. It also ran the largest cosmological simulations to date, providing comparison data for the Large Synoptic Survey Telescope, still under construction.
Summit also ran large computational fluid dynamics simulations. Reported at the AIAA Aviation Forum in June, the FUN3D simulator was used to run a CFD job of 6.5 billion elements and 200,000 time steps. It previously took 2.9 months on a system of 5,000 Xeon Skylake CPU cores. On Summit, using 552 Tesla V100 GPUs (a fraction of what Summit has available), it took four days — a 20-times speed-up. To get the same speed-up with CPUs would take 112,500 Xeon Skylake cores.
Meanwhile, spacecraft computers are finding their way onto atmospheric flyers. In August, NASA’s dual-rotor Mars Helicopter was attached to the belly of the Mars 2020 rover. It is expected to make several test flights in the sparse Martian atmosphere and give operators a perspective of rover surroundings that was not previously possible. The avionics stack includes a Snapdragon flight processor, TI Hercules high-reliability microcontroller and ProASIC rad-tolerant field programmable gate array. In June, NASA selected the Dragonfly rotorcraft to explore Titan, a Saturnian moon with thick atmosphere and low gravity. A radiation-hardened RAD750 computer will control the avionics of this flying laboratory. As with Earth-based drones, a tight control loop of sensors, CPU and actuators must adjust the vehicle’s flight in real time and follow a mission profile. But these flights are after months or years of dormancy in space and targeting different environmental extremes.