Skip to main content
T3 Chips and hardware Cambricon Internship

Cambricon AI Infra internship

Real Cambricon AI Infra internship interview questions covering operator optimization

Operator optimization

Round 1

Operator optimization

  1. In GEMM, why does one thread usually handle an 8x8 compute tile?
  2. What are the main difficulties when implementing operators in CUDA?
  3. Have you implemented prefix sum in CUDA?
  4. What is a bank conflict? How do you reduce bank conflicts?
  5. In Little’s Law, what is the relationship between memory latency and compute latency?

Project experience

  1. Introduce open-source projects you joined.

Round 2

  1. Mainly discussing internship work content and arrangements.