Mixed Precision Objectives

[Awallace3]

Objectives

• [ ] similar to report figure 8, need to compare...:
  • [ ] same convergence (1e-4, 1e-6, 1e-7)
  • [ ] number of iterations to convergence
  • does convergence criteria have number same iterations?
  • Also normalize Wall Clock Time based on number of iterations to get time per iteration for single and double
  • Compare speedup compared to single precision speedup over double (reported 1.92)
• [ ] Recreate Fig. 1 from report with Schulz using single, double, and mixed precision:
  • expect single/multi to lower Schulz by less than half

    Tasks

• [ ] Better single vs. double precision study
• [ ] NERSC recreate environment
• [ ] Mixed precision algorithms:
  • [ ] Precondition?
  • [ ] Convergence switch?
• [ ] DCRD approach with Schulz to avoid A matrix replication costs