Qualcomm Incorporated · NASDAQ:QCOM

Qualcomm — Technical Presentation (April 2022)

13 slides from Qualcomm's technical presentation, April 2022

  1. Slide 1: Qualcomm Cloud AI 100 MLPerf v2.0 Inference Benchmarks

    Qualcomm Qualcomm’ Cloud Al 100 MLPerf ™ v2.0 Inference Benchmarks Qualcomm Cloud Al 100 is a product of Qualcomm Technologies, Inc. and/or its subsidiaries.

  2. Slide 2: Cloud AI 100 MLPerf 2.0 Benchmarks Highlights

    Highligh Continued performance and power efficiency leadership Doubled submissions — 200+ results Expanded sco) partner submissions — Alibaba, Gigabyte, Krai Expanded coverage with Gloria edge gateway platform: Best in class power efficiency and performance * Datacenter BERT Efficiency is 24 best Cl

  3. Slide 3: Most Power Efficient AI Edge Solution — Power Efficiency

    Most Power Efficient Al Edge Solution — Power Efficiency (Queries Per Sec/Watt) Edge power efficiency: ResNet-50 meee 300,3 (DM.2e-15W) 300.3 i234. | (8x Cloud Al 100) 234.1 Qualcomm 199.6 (5x Cloud Al 100) Nvidia (2x A100- PCle -80GB MaxQ)) Nvidia | 56.5 | (x A30, MaxQ) Nvidia Xavier NX (MaxQ) Edge

  4. Slide 4: Datacenter - Power Efficiency

    Datacenter - Power Efficiency ResNetSO SSDLarge BERT-99 (Queries Per Sec /Watt) ID Submitter System QPS/Watt QPS/Watt QPS/Watt 100% Qualcomm 100% Cloud (8x Cloud Al 100) 100% | A1100 100% BERT-99 100% a0. PCI Nvidia oy, 2.0-081 NVIDIA 0 113.4 27 85 (6x A100-PCle-80GB, MaxQ) 57% 80GB, MaxQ 49% BERT-9

  5. Slide 5: Highest peak performance benchmark — MLPerf 2.0 Datacenter

    Highest peak performance benchmark — MLPerf”™ 2.0 Datacenter Closed, Datacenter - Available ResNet-50 offline, peak performance - Qualcomm Cloud Al 100 accelerators Nearest competitor 17% lower with 10x A100 GPU consumes 2x rack space ResNet-50 - Top 5, Closed, Datacenter — Available (Inf/sec) Qualc

  6. Slide 6: Fastest, densest AI Inference Solution in the world

    Fastest, densest Al Inference Solution in the world — MLPerf™ 2.0 Cloud Al 100 servers achieve 3.7x higher rack-level ResNet-50 inference performance than Nvidia A100 servers Qualcomm’ _ Partner Submission 2U = Inspur6U Nvidia Cloud Al 100 platform — platform A100 SXM 80G ile) 20 = 6 8 75W cards — s

  7. Slide 7: Closed Edge Server (Inference/Sec)

    Partner Submission (8x Cloud Al 100) Qualcomm (5x Cloud Al 100) Nvidia (2x A100-PCle-80GB) Nvidia (2x Al00-PCle-80GB) Nvidia (2x Al00-PCle-80GB) Closed Edge Server (Inference/Sec) 118,589 114,869 ResNet-50 ResNet-50 72,835.40 95,832 ResNet-50 98,260 179,618 Cloud Al100

  8. Slide 8: Edge Appliance — (Inference/Sec)

    Edge Appliance — (Inference/Sec) Edge Appliance Vendor (DM.2e 15W) Edge Appliance Vendor (DM.2e 15W) Nvidia (1x A2) FuriosaAl (x Warboy) Nvidia (Jetson Xavier NX Triton) Nvidia (Jetson Xavier NX Triton) 0-106, FuriosaAl 2.0-142 Edge Appliance — (Inference/Sec) Partner Submission (AEDK DM.2 25W) Qual

  9. Slide 9: Edge ResNet-50 Latencies (ms)

    Edge ResNet-50 Latencies (ms) Edge ResNet-50 Latencies (ms) (Lower is better) (Lower is better) mers MS-SS Ratio Nvidia MS.5S Ratio ‘MultiStream (Jetson Xavier NX) MultiStream (x A100-PCle-80GB) SingleStreamn SingleStream Nvidia ike Nvidia J 4.28 | (ix A100-PCle-80GB) (Camemien a) Nvidia wae Furicsc

  10. Slide 10: Cloud AI 100 Closed Datacenter (Inf/sec)

    Cloud Al 100 Closed Datacenter (Inf/sec) ID Submitter System Server Offline Server Offline 2.0123 Qualcomm 18x QAICI00 351,057 8,044, 8,127 2.0-034 Server Vendor 16x QAICIO0 327,119 7.185 7,237 11314 2.025 Qualcomm 8x QAICIO 180,003 181,477 3,567 3,612 5,547 2.0-052. Partner Vendor 5x QAICIO0 112,99

  11. Slide 11: Most Power Efficient - AI Edge Solutions

    Most Power Efficient - Al Edge Solutions Closed Edge Cloud Al 100 ID Submitter System Accelerator Samples/s Watts Gre! Samples/s Watts Gye! Samples/s Watts Gye/ Samples/s Watts Gye! 20-055 FOF’ —AlDevelopmentkit loud ALIOO 225 (4284 5.24 361 3975-909 2.0127 Qualcomm AlDevelopmentkit —_1ud ALIOO 5849

  12. Slide 12: Cloud AI - Closed Edge

    Cloud Al 100 - Closed Edge ID ‘Submitter System Accelerator ieeeren renew tee) te Offline aeons Offline 2.0-055 Partner Vendor Al Development Kit Cloud Al 100 DM.2 25. 17.38 361 2.0127 Qualcomm Al Development Kit Cloud Al 100 DM.2e 15W 084 31 4.63 5,849 16.38 179 2.028 Qualcomm Al Development Kit Cl

  13. Slide 13: ua | CONW\ Nothing in these materials is an offer to sell any of

    ua | CONW\ Nothing in these materials is an offer to sell any of the components cr devices referenced herein. . ©2018-2022 Qualcomm Technologies, Inc. and/or its affiliated Followuson: f ¥ in © @ ‘companies. All Rights Reserved. Qualcomm is a trademark or registered trademark of Qualcomm, Incorporat

Qualcomm — Technical Presentation (April 2022) on 6ix