Workshop on Mixed-Precision Computation

Thursday Oct 23, 2025, 9:30 AM → 3:15 PM America/Chicago

Bldg 201, Rm 102

Esteban Rangel (CPS), Kevin Harms (LCF)

The future generation of computational hardware has added significant computational capability in precisions lower than double precision (64 bit). Modern high-end GPUs have also started to reduce the silicon area dedicated to 64-bit computation, resulting in stagnation or reductions in high-precision performance on a per-GPU basis. In order for high performance computation codes to continue performance improvement on new hardware, application developer must consider the user of lower precision computation. This workshop is an introduction into topic of mixed precision in HPC with presentations on application usage, numerics, and tools.

The workshop is planned to be series of a few workshops with the initial workshop focused within the Argonne computing staff and then growing to larger workshop involving the wider community and vendors.

Mixed_Precision-draft.pdf

1 Welcome & Goals

Speaker: Salman Habib (CPS)

2 Motivation

Speaker: Paul Hovland (Argonne National Laboratory)

3 A Record-and-Replay Workflow for Floating-Point Analysis of GPU Kernels

Speaker: Esteban Rangel (CPS)

4 Half Precision Wave Simulation

Speaker: Longfei Gao (LCF)

10:30 AM Coffee Break

5 Mixed-Precision Quasi-Newton Methods for Optimization

Speaker: Sven Leyffer (Argonne National Laboratory)

6 What Makes Mixed Precision in PETSc Hard? Attempts, Failures, Successes, and Hopes

Speaker: Junchao Zhang (MCS)

7 Roundoff error prediction and correction using CENA and Poseidon

Speaker: Jan Hückelheim (Argonne National Laboratory)

12:00 PM Lunch

8 Group Discussion: Next Generation Architectures

9 Group Discussion: Methods and Tools

10 Workshop Planning and Next Steps

Speaker: Salman Habib (CPS)