Research at Carpathian
We publish what we learn running our own infrastructure and building our own models. Most of it is about doing more with less hardware, because that is the constraint we work under.
What we work on
Our research follows from the way we operate. We run our own hardware in our own buildings, so questions about power, heat, and how much work a machine can do before it needs replacing are not academic for us.
That gives the work a bias we are happy to admit: we are more interested in what runs efficiently on equipment you can already buy than in what runs on hardware that does not exist yet.
- Efficient inferenceRunning capable models on hardware that does not need a purpose-built data center
- InfrastructureCloud architecture and the power, cooling, and scheduling decisions underneath it
- Software practiceHow systems are built, deployed, and kept working once other people depend on them
Current work
Some of this is published, some is still running. Where there is a paper we link it.
Data center architectures that cut power draw without giving up the performance the workload needs. The question we keep coming back to is how much of a facility's energy budget is spent on things the customer never sees.
- Energy
- Sustainability
- Infrastructure
Taking older hardware and making it carry modern cloud workloads. Every machine that keeps working is one that does not become e-waste, and it is cheaper to run than the replacement would have been.
- Hardware
- Sustainability
- E-waste
A byte-level transformer inference engine written in C and architecture-specific assembly, with every internal activation tap-able at no runtime cost. It is the work behind our efficiency claims, and it is open source.
- Inference
- Open source
- Systems
Moving processing closer to where the data is produced, and working out which parts of a workload genuinely benefit from that and which are better left central.
- Edge
- Latency
- Distributed systems
Want to work on this with us?
We work with researchers, academics, and companies who have a problem in this space worth solving. Tell us what you are looking at.