ARM Holdings PLC · NASDAQ:ARM

ARM — Technical Presentation (September 2020)

13 slides from ARM's technical presentation, Sep 2, 2020

1. Slide 1: Initial Performance Results of Nalu on aarch64

   Initial Performance Results of Nalu on aarch64 P3HPC, submission #48 (poster) Srinath Vadlamani, Arm Inc. Sept. 2nd, 2020 online recorded presentation © 2020 Arm Limited (or its affiliates)

2. Slide 2: Motivation: How well does Arm architecture and µ-architecture work with HPC applications?

   Motivation: How well does Arm architecture and µ- architecture work with HPC applications? • A goal we have is to understand a true production code’s behavior on emergent architecture such as aarch64. • Nalu is a SNL production code. • It serves a mission critical role . • It shares common code and

3. Slide 3: Presentation outline

   Presentation outline. • Review ideas for fair performance comparisons. • Review Nalu code and experiment parameters. • Show initials results • µ−architecture (ThunderX2, Graviton 2, x86s) simulation comparisons • Scaling result • Propose plan to diagnose the performance differences across µ-architec

4. Slide 4: Fair performance comparisons sufficient conditions

   Fair performance comparisons sufficient conditions. • Software stack provenance helps the “apples to apples” comparison. • Software versions and build structures should be similar across the µ-architectures. • Also, building and running the same cases (code + initial conditions) across the µ-archite

5. Slide 5: Review of Nalu: Sandia HPC mission critical code

   Review of Nalu: Sandia HPC mission critical code What is Nalu? Experiment Parameters • Low-mach number CFD code from SNL [1]. • Miletsone case meshR2_nt.g • Shares codes exercised in Sierra Framework [2]. • mixture-fraction based turbulent open • In common with Sparc and mission critical jet with a

6. Slide 6: System attributes of HPC systems for performance comparisons

   System attributes of HPC systems for performance comparisons Instance ThunderX2: Stria at AWS M6g.16xlarge AWS M5n.16xlarge: Intel AWS M5a.16xlarge: AMD SNL (Arm) (Arm Graviton2) Threads/core 2 1 2 2 Cores/socket 28 64 16 32 Sockets/node 2 1 2 1 with 4 NUMA regions CPU GHz 2.0 2.5 3.2 2.3 Arch ArmV8

7. Slide 7: Normalized simulation time to m6g

   Newer Arm-based systems show promising performance with baseline case. Single (saturated) node runs of Nalu meshR2_nt.g (2E6 elements). 1.8 Lower is better. 1.6 Normalized simulation time to m6g 1.4 1.2 1 0.8 0.6 0.4 0.2 0 thunderx2: stria m6g.16xlarge: arm m5n.16xlarge: intel m5a.16xlarge: amd epyc

8. Slide 8: Cost per run normalized to m6g

   Aarch64 is advantageous when considering simulation expense Cost for running milestone simulation (single node AWS). Lower is cheaper. 2.00 1.80 Cost per run normalized to m6g 1.60 1.40 1.20 1.00 0.80 0.60 0.40 0.20 0.00 m6g.16xlarge: arm m5n.16xlarge: intel cascade_lake m5a.16xlarge: amd epyc 8 © 2

9. Slide 9: AWS parallel cluster can allow for parallel scaling

   AWS parallel cluster can allow for paralle scaling. 700.00 3.50 NALU: Arm-based Neoverse AWS Graviton2 M6g instances 600.00 3.00 500.00 2.50 400.00 2.00 Simulation time (secs) Speed up 300.00 1.50 200.00 1.00 100.00 0.50 0.00 0.00 0.5 1 2 4 8 Number of compute nodes This AWS parallel cluster instanc

10. Slide 10: Strategies for pursuing performance difference with profiling

    Strategies for pursing performance difference with profiling. • Platform independent tools for profiling the cases can help explain differences across microarchitectures. • Cache misses and computing units usage are to be compared. • Frontend vs. backend stalls can give insight to candidates for bot

11. Slide 11: Conclusions

    Conclusions: • Newer Arm-based systems show promising performance with the baseline Nalu case. • Both in time to solution and cost. • Software management tools such as Spack helps performance analysis across µ- architectures. • The results use same variants thusly reducing the degrees of freedom for

12. Slide 12: References

    References: • [1] Domino, S. "Sierra Low Mach Module: Nalu Theory Manual 1.0", SAND2015-3107W, Sandia National Laboratories Unclassified Unlimited Release (UUR), 2015. https://github.com/NaluCFD/NaluDoc. • [2] H. Edwards, A. Williams, G. Sjaardema, D. Baur, and W. Cochran. Sierra toolkit computation

13. Slide 13: Thank you

    The Arm trademarks featured in this presentation are registered trademarks or trademarks of Arm Limited (or its subsidiaries) in the US and/or elsewhere. All rights reserved. All other marks featured may be trademarks of their respective owners. www.arm.com/company/policies/trademarks Thank you. © 2

Arm — Technical Presentation (September 2020) on 6ix