Control plane for UALink connected AI Accelerators
- Integration with PyTorch and Kubernetes for orchestration
- Complies with latest UALink standard
- Extensible with custom hardware vendor features
Building on our long experience with fabric management, Scalemem is developing a state-of-the-art Pod Controller for UALink. While our main focus is on the Pod Controller itself, we also develop a compliant Switch Management Agent using gNMI/Yang, as well as a Node Management Agent responsible for accelerator discovery and monitoring. We welcome industry collaboration to make sure our software components provide maximum value, both during hardware development and at scale in production.
Our long-term goal is to enable optimally efficient collective operations, accelerated with In-Network Compute (INC) in fault-tolerant UALink virtual Pods.
Full vertical integration
Working the whole stack, from logic gates on an FPGA to orchestrating real PyTorch workloads with Kubernetes, we help hardware vendors showcase their product in a full vertical. Using high-fidelity emulation we develop the full vertical while hardware is in development.