Problem

The bottleneck

As launch costs continue to fall, constellations are growing from dozens of spacecraft to thousands. Yet onboard computing is lagging behind. Most satellite still act as passive sensors merely relaying raw data back to Earth.

Developing applications for space is fundamentally constrained by limited power budgets, memory capacity, and communication bandwidth. Existing compute frameworks and hardware are not designed for these constraints.

Future missions require spacecraft to process information where it is generated, collaborate with neighboring satellites, and make decisions independently. Earth observation, defence and security, autonomous space missions, and space infrastructure all depend on one thing: Reliable Space Native Compute.

Solution

Built for space developers

A unified software framework and heterogenous compute module provides a consistent platform for developing and deploying applications in space. By abstracting hardware complexity behind an easy-to-use platform, developers can focus on application logic instead of implementation details.

SYSTEM MODULE

Architecture

Space Native Compute Platform

01

Mimir Compute Framework

Develop applications using familiar machine learning workflows and compile them into high performing deterministic applications optimized for execution in space.
Retrofit any legacy platforms for immediate performance gains, or unlock modern capabilities on our custom heterogeneous architecture.

02

Mimir Compute Module

A custom heterogeneous compute platform combining CPUs, GPUs, and FPGAs for efficient onboard processing, providing a consistent deployment target for highly performant space applications. 

// models/simple.zig
const mcf = @import("mcf");
const weights = mcf.safetensor();

pub fn model(comptime image: mcf.Tensor) mcf.Tensor {
  const w= weights.get("w");
  const conv = mcf.conv2d(image, w, null, .{ 1, 1 }, null, null);
  return mcf.relu(conv);
}
// main.zig
const Compiler = @import("compiler").Compiler;
const models = @import("models");

pub fn main() !void {
  const graph = Compiler
    .model(models.simple)
    .input(f32, .{ 1, 3, 64, 64 })
    .compile();

  var runtime = try graph.instantiateStatic();

  const image = makeInputImage();
  const output = runtime.run(&image);
}

Mimir Compute Framework

Build once. Deploy anywhere.

Mimir Compute Framework provides a familiar PyTorch-like interface for developing applications. The framework compiles the applications for any LLVM-compatible hardware architecture.

Designed for constrained systems

Instead of relying on heavyweight runtimes and dynamic memory management, the framework compiles applications into small predictable executables built for limited memory, strict power budgets, and heterogeneous hardware.

Mimir Compute Module

Heterogeneous compute

Mimir Compute Module provides custom hardware built specifically for space. It combines different processors and accelerators into an efficient heterogenous compute module.

Optimized for space

Space applications vary widely, and so must the hardware that supports them. Our module provides CPUs for sequential processing, GPUs for parallel tasks, and FPGAs for flexible acceleration of streaming applications.