"LEO Debris Environment Monte Carlo Sim" Requirements:

All versions
- A 64-bit operating system. There are no 32-bit versions.
- Memory: up to about 500 MB of RAM per running task.
- Disk: up to about 200 MB per running task (the checkpoint file grows with the number of objects in the simulation), plus about 3 MB of data files that all tasks share.
- Deadline: 5 days.
- Tasks are run with replication and wingman validation from two matching hosts

CPU versions
- Windows: 64-bit Windows 7 or later.
- Linux (x86-64): static app, no glibc requirement
- Linux on 64-bit ARM (AArch64): static app, no glibc requirement
- macOS on Intel: macOS 10.13 (High Sierra) or later.
- macOS on Apple Silicon (M1 and later): macOS 11 (Big Sur) or later.
- Any x86_64 or aarch64 processor. No special instruction sets such as AVX are needed.
- Each task uses one CPU thread. Run time depends a lot on the scenario: on a current desktop CPU a task averages about 20 to 25 minutes, and so far tasks have taken anywhere from 1 minute to about 6 hours.

NVIDIA GPU versions (CUDA): Windows and Linux on x86-64, and Linux on 64-bit ARM
- Driver: one that supports CUDA 12, that is 527.41 or later on Windows and 525.60.13 or later on Linux.
- GPU memory: 2 GB or more. A task uses up to about 1.3 GB.
- Windows and Linux on x86-64: every NVIDIA GPU from Maxwell (GTX 750 and 750 Ti, GTX 900 series) through Blackwell (RTX 50 series), that is compute capability 5.0 and up.
- Linux x86_64 and aarch64: glibc 2.29 or greater required. Ubuntu 20.04 and newer.
- Not supported: Kepler and older (compute capability 3.x, the GTX 600 and 700 series apart from the 750 and 750 Ti) and NVIDIA's Thor chips.
- Linux on 64-bit ARM: compute capability 6.0 and up with a CUDA 12 driver, for example Jetson Orin with JetPack 6, or ARM servers with NVIDIA GPUs.
- Each GPU task also uses one full CPU thread for its whole run.
- Run time: about 2 minutes on average on a recent card, up to about an hour on slower ones.

AMD GPU versions (HIP): Windows and Linux on x86-64
- Windows: a current AMD Software: Adrenalin Edition (or PRO Edition) driver. It includes the HIP runtime the app uses; nothing else needs installing.
- Linux: glibc 2.29 required. Ubuntu 20.04 and newer.
- Linux: two things must be installed:
(1) OpenCL for your AMD GPU (from ROCm or AMD's driver), which is how BOINC detects the GPU.
(2) The HIP runtime from ROCm 5, 6 or 7 (the library libamdhip64). On Arch and Manjaro the package is hip-runtime-amd; on Ubuntu with AMD's ROCm repository it is rocm-hip-runtime. Without the HIP runtime every task stops right away with "no HIP runtime found".
- GPU memory: 2 GB or more, and the GPU must support double precision.
- Supported GPUs on Windows and Linux:
* Radeon VII, Radeon Pro VII, Instinct MI50 and MI60
* RDNA 1: Radeon RX 5300 to 5700 series, Radeon Pro W5500 and W5700, Pro 5300M to 5600M, Pro V520
* RDNA 2: Radeon RX 6300 to 6950 series, Radeon Pro W6300 to W6800, Pro V620, and the Radeon 660M and 680M integrated GPUs
* RDNA 3: Radeon RX 7400 to 7900 series, Radeon Pro W7500 to W7900, Pro V710, and the Radeon 740M to 780M integrated GPUs
* RDNA 4: Radeon RX 9060 and 9070 series, Radeon AI PRO R9600D and R9700
* the BC-250 board
* Linux only: Instinct MI100, MI210, MI250, MI250X, MI300A, MI300X, MI308X and MI325X
- Not supported: Radeon RX 400 and 500 series, RX Vega 56 and 64, the Radeon 880M and 890M and other RDNA 3.5 integrated GPUs, and older integrated Radeon graphics. If BOINC reports your GPU only as "AMD Radeon Graphics", the server can't tell which GPU it is and sends CPU work instead.
- Each GPU task also uses one full CPU thread for its whole run.
- Run time: about 2 to 3 minutes on average on a recent card.

Choosing CPU or GPU work
If you want only GPU work, or only CPU work, on a computer, turn "Use CPU", "Use NVIDIA GPU" or "Use AMD GPU" on or off in your Kessler@home project preferences. To treat computers differently, put them in different locations (home, work, school) and set the preferences for each location.