Scale-Up Network Architect for Ultra-Low Latency AI Fabrics

Cerebras; United States; Remote; USD 170,000 - 240,000; Full time

Job summary

Cerebras Systems seeks a Network Systems Architect to define scale-up and scale-out network architecture for its platforms, including proprietary accelerators, interconnects, and switching. You will work with application, compiler, runtime, and systems teams to translate workload patterns into fabric requirements.

Your focus is low-latency, scalable fabrics with broad coverage across customer networks, deciding when standards-based or custom solutions are best.

Responsibilities

• Set the multi-generation architecture and roadmap for Cerebras scale-up networks and their interfaces to scale-out and customer-facing networks.
• Work with application, compiler, runtime, and communication-library teams to understand mapping and communication choices, then derive the required bandwidth, latency, ordering, availability, and serviceability.
• Define fabric topology, protocols, and switch behavior, including routing, buffering, flow control, reliability, and fault containment. Connect data-plane choices to end-to-end system behavior.
• Decide when to use standards-based technology or merchant silicon and when a custom protocol, switch, link, or offload is justified.
• Use performance models, traffic simulation, prototypes, and lab data to test architecture choices and set acceptance criteria.
• Write architecture and interface specifications, lead design reviews, and drive cross-layer decisions through implementation, bring-up, and qualification.

Job description

Cerebras Systems seeks a Network Systems Architect to define scale-up and scale-out network architecture for its platforms, including proprietary accelerators, interconnects, and switching. You will work with application, compiler, runtime, and systems teams to translate workload patterns into fabric requirements.

Your focus is low-latency, scalable fabrics with broad coverage across customer networks, deciding when standards-based or custom solutions are best.