Ramanujan-Computing: A Distributed Scientific Computing Platform Built on the World's Idle Devices

Scientific computing is expensive. Researchers at universities pay steep prices for access to supercomputing facilities, while the servers, workstations, and lab machines already on their campuses sit idle for most of the day. Beyond universities, millions of laptops, desktops, and smart devices worldwide are powered on and doing nothing. The Apollo guidance computer — the machine that took humans to the Moon — was only as powerful as a modern scientific calculator. A smartphone today is millions of times more powerful, and most of that power is wasted.

Ramanujan-Computing is an open-source distributed computing platform that changes this. Named in honour of the legendary mathematician Srinivasa Ramanujan, it aggregates idle compute power from any internet-connected device and channels it toward high-value scientific workloads, AI/ML training / inference.

The core insight that sets it apart from existing platforms like BOINC: contributors install the Ramanujan client once, and it can run any computation on the network. Project owners do not write a new client per project — they simply submit their code in the Ramanujan language or Python, and the platform handles the rest. One client. Any computation.

Under the hood, the platform is technically serious. The native interpreter is written in C++, uses a cached-functor execution model with direct memory access, and runs ~14% faster than CPython and ~20× faster than MATLAB/Octave. GPU acceleration is supported via OpenCL. Physics simulations and a full 3.8B-parameter LLM (Microsoft Phi-3-mini) are already running on the platform today.

In this talk, I will introduce Ramanujan-Computing from first principles — the motivation, the architecture, the language, the interpreter design, and the GPU runtime — and show live demonstrations of real workloads running on the platform. I will also talk about the bigger picture: how this approach can help universities unlock the compute power they already own, how it reduces the need for new data centres (and the carbon emissions that come with them), and why compute-donating devices could one day earn carbon credits.

Finally, I will speak to what this project means for open-source contributors. Ramanujan is in its nascent stage — an interpreter being built from scratch, designed to run on any device regardless of architecture, with a long-term vision of becoming what the JVM is for language portability: a universal runtime for distributed scientific computation that can eventually support multiple languages. If you want to build something foundational for the world, this is that opportunity.

1. What Ramanujan-Computing does — and how it differs from BOINC:

Ramanujan-Computing aggregates idle compute from any internet-connected device and directs it toward scientific workloads. Unlike BOINC, where each new project requires contributors to download a new client, Ramanujan uses a universal interpreter — install it once, run any computation. Project owners simply submit code; the network does the rest.

2. The interpreter is genuinely good:

The Ramanujan native interpreter is written in C++ and is ~14% faster than CPython for scientific scripts, and ~20× faster than MATLAB/Octave.

3. How the interpreter works internally — and how it beats CPython:

A deep dive into the internals of both CPython and the Ramanujan interpreter. First, how CPython works under the hood. Then, how Ramanujan approaches the same problem differently. Attendees will leave with a clear mental model of how interpreters work internally.

4. It can use the GPU of idle devices — and run real workloads on them:

Ramanujan supports GPU acceleration via OpenCL. Physics simulations and LLMs like 3.8B-parameter LLM (Phi-3-mini) are already running on the platform. Serious scientific and ML workloads are within reach today.

5. Universities are sitting on untapped compute — Ramanujan unlocks it:

Institutions across India have significant in-house hardware — servers, workstations, lab machines — that sits idle for most of the day, while researchers pay steep prices for supercomputing time they could otherwise avoid. Ramanujan-Computing lets universities put that dormant capacity to work directly. At a national scale, this means growing scientific compute without building new data centres — no new construction, no new emissions. There is also an emerging opportunity: devices that donate compute power could earn carbon credits, creating a tangible incentive for institutions and individuals to participate.

6. What's in it for contributors — this is early, and that's the point:

Ramanujan is in its nascent stage. You have the rare opportunity to contribute to an interpreter being built from scratch — one designed to be more performant than CPython and to run on any device with basic networking and CPU/GPU capability, regardless of architecture. Think of the long-term vision: what the JVM did for language portability — run once, anywhere — Ramanujan aims to do for distributed scientific computation, eventually supporting multiple languages on a single universal runtime.