NPU IP Architect

Project

Work done for a product or program.

• Define MAC array, vector unit and on-chip memory architecture for target workloads.
• Model performance and energy for CNN and transformer workloads.
• Choose supported data types such as INT8, FP8 and INT4 and sparsity schemes.
• Align hardware features with the compiler and runtime team.
• Benchmark against competing NPU IP and customer reference networks.
• Support licensees with configuration, sizing and integration choices.

Function

Work that keeps the discipline running: methods, tools, suppliers and people.

• Track new model trends and update the roadmap.

Project

1. Produce and hand over the role's key deliverables
   • NPU architecture specification
   • Workload performance and energy study
   • Configuration and sizing guides
   • Roadmap inputs
2. Support programsGeneral
   • Agree what each program needs
   • Deliver in time for milestones
3. Manage issuesGeneral
   • Track requests to closure
   • Escalate risks

Function

1. Run the functionGeneral
   • Maintain policies and processes
   • Keep records and controls
2. ImproveGeneral
   • Report performance
   • Lead improvement work