Network So C 산업 — 미래의 정보사회 산업

Network So. C 산업 - “미래의 정보사회: 산업, 기술과 인력” - 2003. 10. 18 임 준서 Paion 1

Agenda I. Demands 시장에 대한 이해의 필요성 Eco-System 의 중요성 – 연못 II. Network So. C 의 기술적 이슈, 경쟁 및 마켓 상황 NP 및 SF 요소기술 소개 경쟁관계 – 별자리도 Map – 역학관계의 파악, 전략적인 방향 결정 2

Agenda III. 변화와 대안 Telco Crash 생태계(Eco-System) 의 변화 변화에 역동적으로 대응하는 실리콘 밸리 기업 들 ; 예) PMC-Sierra, Vitesse, AMCC 자금시장 변화 Network 기술 시장 대안 Network SOC 의 Re-positioning 3

Agenda IV. So. C 산업의 성공적인 미래 Global Trend 중국 시장의 부상 Eco-System 측면에서 시스템 산업, Foundry 제품화 Mind의 기술 개발 - Product Planning Trend 읽기 – 문화 기술 마케팅 – TCP, 시장과 성당 교육적인 측면에서 4

Demand • 시장에 대한 이해의 필요 • 통신 사업자의 서비스 변화에 따른 장비 및 부품 수 요 • Eco-System 의 중요성 5

Core Router : Hype ? Cisco GSR 12416 Juniper M 160 Charlotte Aranea-1 Lucent Nx 64000 Avici TSR Pluris Teraplex $215 M 2002 가을 Aggregated Switching Bandwidth 320~640 G 160 Gbps+ 320 G~1. 9 T 320 G~6. 4 T 400 G ~ 36 T 320 G~19. 2 T 인터넷 트래픽의 증가, 3달에 2배씩 [ 미국 상무성] World. Com 2001년까지, 대형 Terabit 라우터, 대용량 고속 스위치 패브릭, 네 트워크 프로세서 개발에 막대한 투자 6

Investment : Bubble ? 이더넷 신기술 워크샵 2001 -5 -24 7

Big Picture Telco Market Terabit Core Router MPLS Router Metro Core Last Mile Access 802. 11 AP Metro Ethernet (L 3 switch) LAN SW Ethernet IP-DSLAM VDSL Metro Edge Enterprise EPON (OLT) PDSN PSTN BSC Security – VPN, IDS, Firewall EPON (ONU) VDSL Storage Networking – SAN Fabric, TOE Base stations CDN/ Server Load balancing Enterprise Switch 8

KT 의 메트로 및 가입자 스위치 2. 5 G Cisco GSR 12000 백본 라우터 G/E 메트로 코어 대용량 스위치 메트로 코어 중용량 스위치 >36 Gbps, 20 G+128 F 포트 • Riverstone 38000 G/E 1 G Ntopia 스위치 GES 엔터프라이즈 스위치 >120 Gbps, 48 G 포트 • Extreme Black Diamond 8808 1 G VOD server 엑세스 스위치 1 G 1 G DSLAM FE >16 Gbps, 4 G+70 F 포트 • Cisco Catalyst 6506, • Riverstone 8600, • Extreme Alpine 3808 >8 Gbps, 2 G+32 F 포트 • 로커스, 미리넷 • 다산, 코어세스, 텔리언 FE [KT 메트로 이더넷 망 구성도] 9

Broadband Access Carrier overcapacity than demand, i. e. , carriers are in no condition to purchase any equipment. 전반적으로 침체되었던 네트워크 장비 산업은 2002 년 후반기 VDSL, Metro Ethernet 으로 회복되면서, 2002년 도 평균 42% 성장함. KT 등 통신 사업자 주도의 VDSL (Ntopia-V) 및 Metro Ethernet (Ntopia-E) 사업 그리고, Netspot 사업은 Broadband 분야에서 아시아 및 한국의 성공적인 시장 을 창출하였음. 주요 칩셋 벤더의 주된 타깃 시장이 아시아로 이동함. 2003년의 화두는 여전히 VDSL 고도화 였고, 향후 중국 이 최대 시장으로 부상할 것임 국내 중대형 스위치 시스템은 전량 해외 수입 상황임. 10

EPON – 100 Mbps+ Broadband 서비스 , 시스템, 부품 분야에서 한국이 주도할 수 있는 단계임 20 km MTU/MDU EPON-OLT ONU Core VDSL, Ethernet, FTTH Splitter 16: 1~32: 1 Metro Ethernet ONT Home Network Optical First Mile • Minimum fiber/space in CO • No power supply in field • Downstream broadcast (video) DSLAM • 100 Mbps ~1 Gbps Broadband service, cf. VDSL 13~25 Mbps, 50 Mbps 11

네트워크 So. C - 통신 서비스 인프라의 핵심 부품 Backbone & Metro – Telco Market § § Core/Terabit Router - Network Processor, Switch Fabric Metro Ethernet Switch - Network Processor, Switch Fabric Last Mile Access Market § EPON (FTTH) OLT/ONU, VDSL - Ethernet Switch Processor Home Networking § NAT, Home Gateway, Wireless Switch - Low end Network Processor Enterprise Market § § Enterprise Switch - Network Processor, Switch Fabric Security VPN, L 7 Content switch - Network Processor 12

네트워크 SOC 기술의 State-of-the-art • Network Processor • Co-Processor (Classifier, Lookup, Security, Storage) • Switch Fabric • Ethernet Switch Processor 13

Network Processor Value Positioning Network Processor ? § § § A programmable device optimized for packet processing Header parsing, pattern matching, bit-field manipulation, table lookups, packet modification, traffic management Must be done at wire speed, Handle Layer 3 -7 tasks Reduced design cycle for NEV § § § Adaptability to changing standards and protocols being developed– IPv 6, SIP, IPSec, MPLS, Diff. Serv … Faster implementation of Incremental features Performance “headroom” for future upgrades : 50% in usual Total Cost Of Ownership 14

Forwarding and Qo. S functions Ingress Egress Queue scheduling Per Interface Input Classification Ports Rate Limiting Policing/Marking DSCP written L 2/L 3/L 4 filter Drop IP Lookup Switch Fabric Scheduler WRED VOQ DRR, WFQ Output Ports • Usage based Billing Service • Wire-Speed Packet Classification for per-flow rate Limiting • DSCP (Differentiated Service Code Point) written at ingress • Classified and potentially discarded by WRED (Congestion avoidance) • Assigned to the appropriate outgoing queue • Output Scheduling for transmission (priority queuing) 15

Network Processing Platform Control Plane (Protocol Processing & Management ) CPU Data Plane Processing PCI, HT, Rapid. IO, PCI-Ex Network Processor MAC/PHY Gigabit Ethernet/ IEEE 802. 3 ah/ IEEE 802. 11/ VDSL Packet Classifie r Security Engine Packet Processor (Lookup & Forwarding) Switch Fabric Traffic Manager (Qo. S, Scheduling) NPF LA-1 Search Engine - CAM High-Speed Interface Lookup Memory CSIX, SPI 3, SPI 4. 2, NPF SI 16

Function Map • Wire Speed Layer 4~7 Content Inspection • Load Balancing • Per-Flow Classification • Routing (BGP, OSPF) • MPLS Routing • Management (SNMP, RMON, MIB) CPU Network Processor Gigabit Ethernet EPON IEEE 802. 3 ah Packet Classifie r Wireless IEEE 802. 11 Security Engine Packet Processor (Lookup & Forwarding) High-Speed Interface Switch Fabric Traffic Manager (Qo. S, Scheduling) VDSL MAC/Framer, • VPN - IPsec PHY • SSL acceleration • TCP termination • Security Filtering • Security Rule • IP Lookup Search • IPv 6 Lookup Engine - CAM • Layer 3+ Forwarding • NAT • MPLS • Field Manipulation Lookup • Congestion Avoidance Memory • Priority Scheduling • Bandwidth Policing • Service Level Agreement • Billing, Rate Limiting • High speed switching • Non Blocking - VOQ • Reliability/Redundancy • Bandwidth Scalability : 80 G, 320 G, 640 G 17

Network Processor Issues Processing Bandwidth l l l 1 G (1. 488 Mpps) OC-48 (7 Mpps) OC-192 (32 Mpps) 40 G (120+ Mpps) Deep packet classification, 1 M LUT Multi-Processors : 100%+ compute headroom More powerful engines – VLIW, SMP, Dynamic multithreading Extensive use of hard-wired co-processor, classification, etc Memory Bandwidth § Layer 3 10 G requires 5 GBps + sustained, 6~10 GBps peak l l l § 4 * Rambus = 6. 4 GBps peak 256 bit DDR-266 = 8. 5 GBps peak 300 MHz DDR = 9. 4 GBps peak Layer 7 10 G l l Need 20 GBps sustained memory BW just for packet queuing Combination of embedded DRAM (bandwidth) and external memory(capacity) Line/Fabric Interface § § § Need higher bandwidth & low pin counts CSIX-L 1 scales poorly to 10 G - Too many pins SPI 4. 2 standard for 10 G line side 18

[참고] Processing Engine – Old Data Core Forwarding Clock rate Process Power Intel IXP 6 packet engine(Strong. Arm) 3. 3 Mpps 232 MHz 0. 28 um 4. 3 W AMCC 7120 2 packet engine (n. Pcore RISC) 3 Mpps 220 MHz 0. 18 um 4 W IBM Rainier 16 pico engine (32 b RISC) 16. 6 Mpps 133 MHz 0. 18 um 12 W Cport C-5 16 channel processor (MIPS-1) 16. 7 Mpps 200 MHz 0. 18 um 15 W Agere 6 VLIW+2 FPP L 3 OC-48 133 MHz 0. 18 um 2*12 W AMCC 7120 core 220 MHz 0. 18 um NA EZchip NP 1 64 packet engine 200 MHz 0. 13 um L 7 OC-192 19

[참고] Bandwidth – Old Data Line BW Fabric BW Memory BW Next Step Intel IXP 2. 6~3. 3 Gbps IX 2. 6~3. 3 Gbps 800 MBps (64 bit SDRAM 256 M) 400 MBps (32 bit 18 M SRAM 100) Move to standard interface SPI 4/CSIX instead of IX AMCC 7120 5 Gbps Vi. X 5 Gbps 440 MBps (32 bit SRAM lookup) IBM Rainier 8 Gbps DASL 2 x 7. 5 Gbps 1. 6 GBps (256 M DDR ) 800 MBps(SRAM) Uni. Link C-port C-5 5 Gbps CSIX 6. 4 Gbps 2 GBps (64 bit 128 M SDRAM), 1. 1 GBps(16 M SRAM-133 MHz) Follow standard interface Agere 5 Gbps Utopia 5 Gbps 2. 2 GBps (64 bit SDRAM), 1. 06 GBps(SRAM) Standard fabric interface SPI 4/CSIX AMCC 7250 5 Gbps Vi. X 5 Gbps 2~4 GBps Vi. X to SPI 4 conversion chip, Next integration EZchip NP 1 20 Gbps CSIX 25 Gbps+ >60 GBps (5 MB on-chip DRAM) CSIX and SPI 4. 2 20

[참고] Interface Utopia (ATM Interface) § § § Level 1 : 8 bit, 25 MHz OC-3(155 Mbps) Level 2 : 16 bit, 50 MHz OC-12(640 Mbps) Level 3 : 32 bit, 104 MHz OC-48(3. 2 Gbps) OIF SPI-3 SPI-4. 1 : 64 bit, 200 MHz OC-192(10 Gbps) SPI-4. 2 : 16 bit, 622 MHz OC-192(10 Gbps), LVDS 311 MHz DDR CSIX L 1 § CFrame : 32 xn bit(n=1, 2, 3, 4) with 100/250 MHz up to 32 Gbps l l l § § LVCMOS(166 MHz), LHSTL(250 MHz), Trace Length : 6~8 inch Header(2 B) – Header Ext(4 B) – Payload(256 B) – Parity(2 B) l § § 32 bit @100 MHz 32 Gbps, OC-48 64 bit @200 MHz 12. 8 Gbps, OC-192 128 bit @100 MHz 12. 8 Gbps, OC-192 Class (8 bit 256 Class) , Destination (12 bit 4096 CSIX port) Xon/Xoff port-level flow control, In-band control CFrame Too many implementation option, Not optimized for 10 G NPF SI : § § § Logical protocol scalable from 10 Gbps to 40 Gbps 16 bit LVDS datapath 622 MHz NP-NP : 256 ~1 M port, 256 class, NP-Fabric : 64 K port, 256 class 21

t. CAM Lookup Processing Search entries that match key(comparand) § § § CAM returns address of the matching location Most designs optimized for 72 bit search, 72 -bit comparand bus Often runs at 2 xcore clock or in DDR mode 144 bit key Write operation in single memory cycle for fivetuple lookup 576 bit search for IPv 6 Performance 2 Mb 66~100 Msps $70 4. 5 Mb 100 Msps $150 0. 18 um 0. 15 um 9 Mb 133 Msps $300 0. 13 um 22

Flow Classification for Qo. S Classification (classify flows) Rate Limiting (Per-flow RL) Scheduling (Priority) Control packets Voice flows HTTP flows FTP flows 23

Per-Flow Rate Limiting Flow § § A sequence of packets with one common characteristics based on any field at certain layer for some period of time Associate IP packets to flows, provide guarantees for specific flows Rate Limiting § § § Bandwidth allocation for each Customer/Application/Flow If current rate is within agreed controlled(in-profile), admit Otherwise (out-of-profile) discard or downgrade packet Bandwidth Provisioning granularity Packet Classification § § § Identifies the flow to the which an arriving packet belongs. Identification may require complex operations Stores this information internally once, then other blocks of NP operation for each packet using flow ID 24

NPU Industry System vendors want Complete § § Solutions NPU, search engine, MAC, PHY, fabric l Reduces TCO, project overhead, and design time l Easier for new OEMs to enter market l Fast market creation using off-the-shelf NPUs Successful vendors must offer all parts – driven continued M&A Big CPU vendors jumped in § Intel, IBM, Motorola PHY vendors moving upstream § § AMCC, Vitesse, Agere Software is a new challenge Related players acquired l Switch. On, Orologic, Maker, Net. Boost, Soft. Com, Trillium, QED, Si. Byte Startup continue to emerge l EZchip, Clear. Water, Cognigine, Bay Micro, Internet Machine, … 25

Competition Map [NP 1] AMCC (Market 40%) 3 rd Generation n. P 7250 OC-48 Sample June 2001 n. P 3400 2 G+24 F n. P 7120 2 x. GE 2000 n. P 3404 4 x. GE May 2001 Integrated Switch 2000 NP Pioneer AMCC 7250 w. Same 7120 core w 220 MHz, 0. 18 um w 2 SPI 3 (2 GMII or OC 48 c) Duplex w 10 Gbps TM (5700) on Line card w. Vi. X 5 Gbps (no CSIX) w 2~4 Gbps (Memory BW) w$300 AMCC 7120 4 th Generation n. P 7510 10 G 2002. May n. P 7570 10 G 2003 n. P 7270 OC-48 0. 13 um 2 packet engine(n. Pcore RISC) Q 2 2002 220 MHz, 0. 18 um n. P 3450 3 Mpps, 5 Gbps (line rate) 2 G+24 F Vi. X 5 Gbps (Fabric interface) Sample 440 MBps (32 bit SRAM lookup) Strong Compiler, Design Winner 3 Q 03 AMCC n. P 7510 • 0. 13 um Full-Customized n. Pcore • 10 G (Simplex) Sample 4 Q 01 • Six engine @ 330 MHz • Vi. X/SPI 4 conversion chip • More Headroom • Ingress/Egress 2 K Flow each AMCC n. P 7570 n. P 3454 4 x. GE • SPI 4. 2 Duplex version • Biggest Competitor IXP 2800 Sample 3 Q 03 INTEL ( 11%) IXP 1200 • Strong. Arm, 3 Mpps@200 MHz • IXBUS (External MAC) • 3 rd software – Teja, IPinfusion, Lynux. Works • SDKv 2. 0 May/2001 (free) • C compiler $2000/site • Aggressive design win Push IXP 2400 • Simplex OC-48 • 7 Mpps @600 MHz • CSIXL 1 over SPI 3 • 0. 18 um, Xscale based • $200~300, Sep. 2002 IXP 2800 IXP 2850 • 10 G (Simplex) • 2800 + Security • 30 Mpps with 16 packet engine(1. 4 GHz) • 700 MHz Xscale at 0. 13 um • CSIX over SPI 4. 2 • >$400, Apr. 2003 26

Competition Map [NP 2] IBM (19%) Agere ( 9%) NP 4 GS 3 • 4 GMII with CSIX/UDASL • 0. 13 um • 4 Gb. E MAC • Power. PC 405 + 16 packet engine • Dropped • 8 Gbps line, 2 x 7. 5 Gbps fabric (DASL) • Difficult to program - No compiler, support customer base • Limited performance and poor scalability • 0. 18 um, $400 • 2001/Q 1 Production NP 4 GS 3(Rainier) Motorola (15%) C-5 • 4 Gb. E, OC-48 • 32 bit Utopia, CSIX • MIPS-based compiler • 233 MHz 0. 18 um • Production 2001, $400~500 C-5 e APP 550 Payload NP 4 GX Payload. Plus APP 550 • Duplex(OC-48, 4 Gb. E) • 5 Gbps, 3 chip Solution ($585) • TSMC 0. 13 um • High-end ATM applications • 266 MHz • On-chip SAR (OC-48 speed) • $550, 2 Q 03 Volume • Scale well OC-192 • Traffic Manager $550 • TSMC 0. 18 um Full custom, Late/2001 APP 520 APP 750 • Simplex (10 Gbps) • SPI 4. 2 • Connection to PI 40 • Agere 0. 16 um • Sample 2003 • NP+TM (860+$700) • Duplex (2 Gb. E, OC-48) • $195, Apr. 2003 C-3 e 4 Gb. E, OC-48 1 Gb. E or 8 FE • 32 bit CSIX • 16 bit Utopia • 266 MHz 0. 15 um • 180 MHz 0. 15 um • Production 2002, $450 • Production 2002, $200 • Traffic Manager Q-5 ($575) • Traffic Manager Q-3 ($235) C-10 10 Gbps (Simplex) • 600 MHz 0. 13 um Custom design • SPI 4. 2 • Sample 2 H 03 with Q-10 27

NPU Industry Update NPU Market Forecasting § § NPU Market $1. 5 B @2005 [Cahners In-Stat, 2002] NPU Market $728 M @2005 [Gartner Dataquest, 2002] NPU Market 2002 § § § $65 million (total Revenue) - A far cry from the hundreds of million dollars that IC maker till believe is possible 600 design wins @2001 leads to $65 million @2002 Market is questionable with annual $50 million spending NPU Player § § § § AMCC market leader : 38 % Intel : 200 design wins Motorola : 15% IBM : 19%, High-end, but pulled out Agere : Late player in IP, but actively Vitesse : pulled out from long-haul business Many startup ceased operation, even Big player pulled out Key Issues § § What is sweet-spot application? Storage, Security, L 7 for IDC Who will be remained? Alternative? 10 G NPU, Access, Wireless 28

Constellation 40 G NPU 10 G+ NPU Co-processor Acorn Solidum Silicon Access EZchip NP-1 Fast-Chip Intel IXP 2800 Switch. On AMCC n. P 7510 Lexra CAM Net. Logic Cypress/Lara Motorola C-5 Broad. Com Agere SB 12500 APP 550 1 G NPU Mind. Speed CX 27470 AMCC n. P 7120 Intel IXP 1200 EZchip NP-2 Xelerated Packet Cognigine Bay Micro Paion GEP 2 F 24 IBM Rainier Siber. Core Internet Machine Clearwater Agere APP 750 2. 5 G NPU IDT MUSIC Clear. Speed AMCC n. P 7250 Paion GEP 2 C 02 Intel IXP 2400 Storage processor (TOE, i. SCSI, FC) Trebia Silverback Alacritech Adaptec Aristos Qlogic Vitesse IQ 2000 Wintegra Switch. Core CXE-16 Galileo Gal. Net-3 Allayer AL 1032 29

[참고] Market Size and Market Share [Dataquest, 2002] 30

t. CAM Lookup Processor n IDT : Current market Leader n n Cypress/Lara : 2 nd leader n n n 18 Mb, 9 Mb+LA-1, 288 bit search with 50 Msps, $250, Sampling [2002. 8] Ultra 18 Mb Sampling [4 Q 02] KLSI : Early lead with 4. 5 Mb n n n 18 Mb NSE Sample Second-source agreement with Micron[2003. 1] Micron pulled out [2003. 3] Siber. Core : superior n n NSE 1000/2000 - 18 Mb & LA-1, 233 Msps, Sample [4 Q 02] Leverage Ayama technology M&A [Feb. 2002] Net. Logic : aggressive performance n n Major wins for Cisco, Acquiring IBM solution 18 Mb & LA-1, 100 Msps, TSMC 0. 13 um, Sampling [2002. 9] 250 Msps version 18 Mb Sample Planned [1 Q 03] 9 Mb delaying with competitors shipping 18 Mb Retreat from t. CAM market at Jan/2003 Micron/Music : Losing ground n n n Acquisition of Music - for low-end t. CAM, 2002 Abandon internal development, Partnership with Net. Logic But, Micron pulled out from t. CAM market [March 2003] 31

Packet Classifier L 4~7 Classification Chip Application for Content Switching, Firewall, Virus scanning, Intrusion Detection Rule memory PMC Classi. PI 2. 5 Gbps On chip (16 K) Layer 7 at lower rates, Based on Switch. On IDT/ Solidum PAX. port 2. 5 Gbps 3 x 64 bit ZBT SRAM First, but limited appeal 2. 5 Gbps (SPI 3, 2 Gb. E) TSMC 0. 18 um, June 2002 Sampling, $199 Solidum M&A (2002. 9, $10 M) Fast-Chip Policy Edge 10 Gbps On chip (16 K) Powerful, Not Yet Sample 2001. 10, $450 Closed Operation Hifn MFA 4 Gbps PC 133 SDRAM Late Player 0. 18 um, $350, Sampling 2 Q 02 M&A (2000. 5, Appititue) Safe. Net Re. GXP 2 G 2. 5 Gbps On Chip (1 MB), Better than PMC, Competing IDT 8 K rule 2. 5 Gbps (SPI 4) Sampling 4 Q 02, M&A (Raquia, 2003. 3, $7. 5 M) 32

[참고] Security Engines Encryption Engines best suited for IPsec VPN applications Integrated high-speed encryption engine with Network Processor will drive future Gbps IPSec, SSL acceleration Broadcom : § § § BCM 5850 current fastest IPsec chip 4. 8 Gbps IPsec (3 DES, SHA-1, AES) PL 3 interface to BCM 1250 processor 2002 Oct Sample, $520 Move toward integrating security chip into other products § § § First flow-through 2~10 Gbps IPsec & SSL performance 5 Gbps IPsec, 38, 000 SSL 2002. 4 Sampling, $295~$795 § § § Market leader in security, but lost the technology leader to Cavium, Corrent, and Layer N Single chip (Encryption, Classification) 20 Gbps IPsec at 10 Gb. E Flow-through SPI 4. 2, HT, PCI-X Sample Early 2003 Merging Appititude, Meter. Flow, into Security processor Cavium Nitrox. II Hi. Fn HIPP-3 l Deep packet analysis after packets are decrypted § 10, 000 or more SSL handshaking per second § § § 10 Gbps On-chip encryption engine with $245 premium SHA, 3 DS, AES for IPsec, SSL application General Sample 2 Q 03 Intel 33

Switch Fabric Ingress Queue(VOQ) 1 1 Crossbar Egress Queue(OQ) 1 1 2 N Grant 2. 3. 4. 5. 3 N 1 N Configuration Request 1. 2 N N 1 3 1 Uni/Multicast Arbitration Flow Control Scheduler Ingress data when there is no backpressure signal, i. e. , queue is not full Each ingress Q requests for cell transfer to Egress port Ingress Request Generation Block generates and send requests to destination ports Grant Generation Block selects one request from ingress ports and grant a transfer Start cell transfer 34

Switch Fabric Requirement Switching Bandwidth Increase § § Single stage 40~80 G with 16 or 32 ports (Today) Need 320 G+, Terabit Scalability required Shared memory : 80 Gbps physical limit? Crossbar : High speed implementation, intelligent scheduling Efficient Scalability § § Multi-stage architecture is inevitable for 10 s of Tbps router - bit-slicing method is limited Performance is a factor of flow control, priority schemes, latency and Non-blocking behavior for Unicast/Multicast/ Broadcast traffic Total Cost of Ownership § § Number of chips, power consumption, Multi shelf interconnect Low Power high-speed Serial Technology N+K redundancy trade-off is driven by system cost Support for multiple services § § IP, Ethernet, ATM, TDM 35

[참고] Market Forecast [2000] 36

Switch Fabric Market Update [2003] Market Various analysts projected it to exceed $ 1 billion by 2004, with CAGR 60% § Market was too much over forecasted [Gartner Dataquest, 2000] Service Provider Router Market [Synergy Research Group, 2003] § But as we enter 2003, telecom continues to be moribund § New chassis design launches are important for switch fabric market; switch fabric is one of the first architecture decision point § 37

Fabric Trend Update More system upgrades than new chassis § § Cap-ex spending is down Design-win momentum slowing Long-term fabric market is healthy § § Most OEMs using merchant fabric silicon from in-house ASIC New design will start again on 2004 for new box (06) l § Backbone product design cycle : 6~7 years, last cycle was ’ 99 Volume and prices continue to decline Slowdown forcing consolidation § Alliance are critical for success Optimize cost/performance curve § § Integrated fabric and Traffic Manager Focus shifted from 10 G to 1~2. 5 G application Trend toward standard interface § § § Standard CSIX/SPI allows flexible choice of NPU CSIX-L 1 for OC-48, Many options in the spec NPF-SI addressed 10 G applications 38

Switch Fabric Players § PMC-Sierra l l § IBM l l l § Through acquisition, AMCC provides total solutions Latest product, n. PX 5800, scales to 160 Gbps being cancelled Power. PRS will replace 5800 Yuni networks, fills the 640 Gbps fabric solution Agere l l l § Leading incumbent vendors Giga. Stream offers leading integration and power dissipation at OC-48 Tera. Stream has less competitive feature sets and lags in throughput AMCC l l § 10 Gbps to 320 Gbps, So far leader in fabric market Power. PRS fabric marketed together with Power. NP network Processor Pulled out Fabric market, Sold to AMCC at $47 million Vitesse l l l § Pulled out Fabric and network processing business But the ace up MPC-Sierra’s sleeve could be microprocessor division Third generation fabric Pi 40, first commercial multi-terabit switch fabric Leading product at throughput of 320 Gbps and above With 10 Gbps NPU, Agere offers all the major components of router designs Startups l l l Mindspeed, Zetta. Com, Erlang, Paion Many a startups pulled out Power. X, Trans. Warp, Peta. Switch, Tau, Tera. Blaze, Tera. Cross 39

Competition Map [Switch Fabric] Agere PI 40 80 G/2. 5 Tbps Marvell • Prestera FX 9130 Tera. Chip • 80 G/5 Tbps • 160 G/1. 28 Tbps • 2003 • 80 Gbps(1 SF+4 QM) • 2003 AMCC • 2. 5 Tbps(32 SF+128 QM Tera. Cross ) Peta. Switch • Glimps 1000 n. PX 8005 • Pisces 40/1. 2 Tbps • 2 Q 02 Sample • 80 G/1. 2 Tbps • 320 Gbps(8 SF+ 2 SC+QM) • Agere 0. 16 um, $1100 • 320 Gbps(3 chip) • Intel FAB + IXP 2800 • Sample 1 H 02 • 640 Gbps(16 SF+ 4 SC+QM) • UMC 0. 15 um IBM • 1. 2 Tbps(32 SF+8 SC+QM) 2. 5 Tbps 1. 2 Tbps PMC TTX 640 Gbps 320 Gbps PMC ETT 1 320 Gbps IBM Q 64 G 160/320 Gbps AMCC Npx 5800 20/160 Gbps EQ+ • 160 Gbps(4 SF+16 QM) • 320 Gbps(8 SF+32 QM) • 1 H/02 • 20 Gbps(1 SF+ 1 QM) • 160 Gbps(8 SF+ 4 QM) 80 Gbps Agere PI-X 20/80 Gbps IBM 64 G 40/80 Gbps • 40 Gbps(1 SF+ 4 QM) • 80 Gbps(2 SF+ 8 QM) AMCC 5500 20/40 Gbps Vitesse CS 32/64 Gbps Vitesse Giga. Stream 20/80 Gbps Mindspeed i. Scale 40/320 Gbps Vitesse Tera. Stream 80/160 Gbps Broadcom Zetta. Com Internet TAU Machine • 40/640 G • Gemini • 40 G(1+4) • 40/640 Gbps • 320 G(8+32) • 80/640 G • 3 Q 02 Erlang • 640 G(16+64 ) • 80/640 G • 3 Q 02 • 40 Gbps(1 SF+2 QM) • 320 Gbps(8 SF+16 QM) • TSMC 0. 18 um • 80 Gbps(2 SF+8 QM) • 160 Gbps(4 SF+16 QM) • 2 Q/02, CSIX, 16 class • $1154+$690 • 1 Q/02, Utopia, 8 class • $240+$400 • 2 Q/01 2002 2003 40

Switch Fabric Class 분야 가입자 및 에지 메트로 코어 대역폭 20 ~ 80 Gbps 40 ~ 320 Gbps 80 ~ 640+ Gbps 칩 가격 경쟁력 집적도, 확장성 고성능, 대용량 포트, 전 력소모, 확장성 문제점 AMCC n. PX 5800 • 20 Gbps 주요 경 • 160 G (8 X, 4 Q) 쟁사 Vitesse Giga. Stream • 20 Gbps • 80 G (4 X, 8 Q) Vitesse Tera. Stream • 40 Gbps • 160 G (4 X, 16 Q) Zetta. Com IXS 250 • 40 Gbps • 640 G (16 X, 64 Q) IBM PRS-64 G, Q 64 G • 40 Gbps • 320 G (8 X, 32 Q) AMCC n. PX 8005 • 40 Gbps • 1. 28 T (32 X, 8+128*2 Q) Mind. Speed i. Scale • 40 Gbps • 320 G (8 X, 16 Q) Agere PI 40 • 80 Gbps • 2. 5 T (32 X, 128 Q) 41

[참고] Fabric Architecture Comparison Product Link speed Switchin g Bandwid th Aggregate Bandwidth Switch Architecture Cell/Frame Interface IBM PRS 64 G 32 x 32@O C-48 128 Gbps 256 Gbps (2 chip) Single stage shared memory VOQ , 4 level priority Fixed/Variabl e (80 B~160 B) DASL 500 MHz 40~80 Gb ps 320 Gbps (4 chip) 4 class, SP, VOQ(1024 cell, 40 KB), OQ(512 cell), 4 shared memory bank(1024 row) 64 -80 B Uni. Link 2. 5 GHz IBM PRSEQ AMCC n. Px 5800 16 x 16@10 Gbps 40 Gbps 320 Gbps (8 module) Shared memory, VOQ 256 Queue, 4 class, Strict Priority 64 B Packet/Cell Vi. X v 3 (VIX to SPI-4 phase 2 conversion) AMCC n. Px 8000 32 x 32@2. 5 Gbps 320 Gbps 1. 2 Tbps Crossbar, 1024 VOQ (128 port, 8 class), , Two * i. SLIP, Channelized Max. Matching Fixed cell (64 / 80 B ) Ser. Des Vitesse VSC 882 32 x 32@O C-48 80 Gbps 320 Gbps (4 chip) Crossbar, 16 class, VOQ (32+1, 2 Mb), OQ(256 kb), 80 B Variable/Fixe d CSIX (plan) Agere PI 40 32 x 32@2. 5 G, 64 x 64@2. 5 G 80 G, 160 G 320 G (20 chip), 2. 5 T (160 chip) Shared memory+Crossbar 3 stage Clos, VOQ – 8 K cell(UC), 1 K Variable/Fixe d (64, 72, 80 B) 42

변화와 대안 • Retrospect • Telco Crash • 생태계(Eco-System) 의 변화 • 변화에 역동적으로 대응하는 실리콘 밸리 기업들 • 자금시장 변화 • Network 기술 시장 대안 • Network SOC 의 Re-positioning 43

Retrospect Intelligence (Layer) 7 6 10 G+ NPU EZchip NP-1 Intel Lexra Agere AMCC n. P 7510 Internet Machines Silicon Access Clearwater Cognigine Bay Micro 2. 5 G NPU AMCC n. P 7250 Vitesse IQ 2200 Mind. Speed CX 27470 Broad. Com SB 12500 5 1 G NPU Xelerated Packet Device EZchip NP-2 Clear. Speed Multi-Service Platform Next Generation Mobile Infrastructure Giga/Terabit Backbone Intelligent and Secure 8 4 Motorola C-5 IBM Rainier Vitesse IQ 2000 4 Network SOC Roadmap 32 9 Cisco Juniper Charlotte Lucent Avici Pluris GSR 12416 M 160 Arenea-1 NX 64000 TSR Teraplex 56 Nexsi Alteon Arrowpoint Foundry F 5 BGP 8000 180+ CSS 1100 Server. Iron 800 IP Intel Custom Web. IC ASIC MMC Iron. Core ASIC Metro Ethernet Broadband Foundry Big Iron 15000 Fast Iron III GEP 4 C 04 Intel IXP 1200 AMCC n. P 7120 3 Galileo Gal. Net-3 Intel Switch. Core CXE-16 Intel Broad. Com Acute Allayer AL 1032 24 Foundry Fast Iron II 32 Foundry Fast Iron II+ 64 96 48 32 Extreme Alpine 3808 Summit 7 i PMC-sierra ETT 1 AMCC n. P 5800 Power X SX-16 Zetta. Com ZEST-IXS 4 8 16 32 120 139 144 160 256 Cajun Cisco 40 60 Extreme Black Catalyst Extreme Black Diamond 6816 P 882 6513 7606 Riverstone Diamond 6808 RS 8600/ RS 16000 64 G SF 2 In-House ASIC Riverstone RS 38000 320 G SF IBM PRS 64 G Mind. Speed CX 27302 Paion GES 0032 64 Agere PI 40 X/C PMC-sierra TTX IBM PRS-EQ Vitesse VSC 882 128 640 G SF AMCC n. P 8000 Velio VC 2002 Paion MOSAIC 256 512 Bandwidth (Gbps) 2001. 10. 17 전략기술개발사업 기술기획 44

45

Telecom Collapse Largest Network equipment buyers Capital spending cut from $116 B(00) to $46 B(02) : 60% drop Telco revenue $197 B for voice, $63 B for data [2001 US] Carriers spend 15% of revenues on equipment Cut off 500, 000 jobs in two years 61 service provider bankruptcies 11 of the top 25 money losing firms in 2001 were telecom § § Total $107 B losses $470 B of debt $ Not going to return to the late 1990 s in a near-term recovery New business strategies are required 46

Impact Effects on Vendor § § DSL, wireless, cable vendors devastated – customers collapsed Components awful – no new design wins Optical very bad – revenues off 80% Core routing bad – revenues off 50% Effects on Financial Market § 60, 000 jobs lost in investment banking since 1/2001 - no acquisitions or IPOs § 2, 500 companies were acquired or went public in the 1996 -2000 boom Over 10, 000 venture-backed private companies now At least 7, 500 (75% or more) will run out of money in 2002 -2005 § § 47

Intel AMCC EZchip Wintegra 생태계의 변화 Agere Motorola Silicon Access Zetta. Com Internet Machines Broadcom Marvell Vitesse Zarlink Power X 48

Market Value Erosion Leader Market cap (B) Market cap 9/30/00 (B) 9/30/01 Market cap (B) 9/30/02 Chip BRCM+PMCS+ AMCC $114 $9 $6. 5 Optical JDSU+GLW Network Vendor JNPR+CIEN+SCM R+SONS+RBAK+ FDRY+EXTR $177 $158 $17 $10 $6. 2 $7. 5 Major CSCO+LU+NT $677 $127 $105 ILEC BLS+Q+SBC+VZ $452 $1578 $407 $569 $209 $335 Total 49

Stock Market 50

Alternative Find the bottom Look for a safe niche Hibernate and wait for an upturn Seek government intervention Explore non-US opportunities Proper market & technology § § Vendors/investors forced to abandon innovative nextgeneration products Telco spend money to save money, not for new services Go after enterprise market instead? What about consumer/residential market? 51

Alternative Technology. . Still Most Promising § § § Security (SSL/VPNs, firewalls, IDS, identity management, attack mitigation) Wireless (WLANs, roaming, fixed broadband) Storage networks (NAS, SANs migration to Storage networks over IP) Still Must § § § Switching (metro, more VPN “flavors”) Sell to technologically-backward incumbents New SONET and MSPP , softswitches Lagging § § § Vo. IP (SIP, ENUM, virtual PBX services) Transport (10 Gig Ethernet) Optical (40 Gig SONET/SDH, optical transmission, GMPLS) 52

Example: Fusion of Home Network and Entertainment Cluster With broadband growth starting to slow, Telco’s are serious about investigating the smart home network segment for value added broadband services -- Residential gateway Integrates cable modem/router/wireless access point § § § Home networking has become a bigger focus of the CE industry Products are emerging to enable connectivity between the home data network and the entertainment cluster at reasonable price point The next wave of interconnectedness between devices will not for basic data/broadband sharing, but instead for the fusion of the PC and entertainment worlds in the form of media center Silicon vendors rush § § Standard Activity : 802. 15 Microsoft – the biggest splash with the media center edition of XP Intel – media center reference design and the push to get CE vendors to create a new product category around the media center Motorola - connected home 53

Example: Home Networking Disk Array (Server) UWB Wi. Fi 1394 UWB IEEE 1394 802. 11 CPU Switching FE PHY Optional IPSec SA and SSL Content memory EPON(FTTH) Super VDSL Metro Ethernet PCI NPU 802. 11 AP Home Media Server Firewall Security Billing Ratelimiting EPON MAC 8 x FE System memory 54

Example: Wireless+Security Qo. S with guaranteed SLA, Bandwidth rate limiting § Application-based flow classification for Vo. IP and streaming media § AP & L 3/L 4 Load Balancing Wireless Mobility Security § § § Encryption for WEP, TKIP and AES, Kerberos Security VPN(L 2 TP/IPSec) Wireless Management Serial and Power over Ethernet 802. 11 AP 63 m 55

Example: Wireless/Mobile Coexistence HLR/VLR IWF PSTN Radio Access Network E 1/T 1 Voice Network MSC: Mobile Switching Center BTS 1 E 1/T 1 AAA HA OAM E 1/T 1 BTS n Roaming / Handoff ATM/IP BSC Data Network Fast Ethernet Base Station Controller PDSN Packet Data Serving Node Data Network Internet Wireless LAN 802. 11 AP Metro Ethernet Switch Wireless Switch 56

Repositioning: Ethernet Switch Processor Market § § Gb. E switch silicon revenue : $348 million in 2002, $540 million 2004, and $1. 063 billion in 2007, 25% CAGR, [IDC, March 2003] Broadcom and Marvell owned 90 percent of the market in 2002, the split was about 6040, Mostly Fast Ethernet market Sweet-spot : 24/16/8 port Until now, Gb. E switches were considered to be too expensive for all but enterprise applications. However, Gb. E-enabled laptops, PCs and network interface cards (NICs) are driving the cost of Gb. E to all-time lows. This has accelerated the demand for higher bandwidth Gb. E switch performance across all segments of the market, further increasing volumes and reducing cost. Player § § Broadcom – announced 24 port and 16 port Gb. E switch on Sep. 2003, and provides 4 -, 5 -, and 8 -port Gb. E switch with integrated PHY, General sampling Oct. 2003 Marvell – Announced 24 port Gb. E switch, Prestera-EX 242, on Feb. , Integrated SERDES with 24 port Gb. E on Apr. 2003, Vitesse – announced Stapleford(VSC 7303) – 24 port Gb. E switch chip, doing Layer 4 processing on April. 2003, Announced Heathrow-II, 16 port Gb. E on 2002 Spring Others : Zarlink, Realtek, ADMtek, Switch. Core, Via, Paion Issues § § Integration up to Layer 3, Filtering. SERDES and PHY consideration in future Cost : < $6~8 / Gb. E port Power : < 0. 2 ~0. 25 watt/ Gb. E port Foundry : 0. 13 um CMOS process 57

Competition Map [Ethernet Switch Processor] 40/160 Gbps Tera. Stream 24 GE VSC 7303 (Stapleford) 20/80 Gbps Giga. Stream 16 GE VSC 7301 (Heathrow-II) VITESSE Cross. Stream M&A(Sitera) M&A(Exbit) Drop(NPU) MARVELL 24 GE+ 80 G~5 Tbps 24 GE Prestera EX 242 $200+ 4 G+48 FE 40~640 Gbps M&A(Galileo) 13 GE 2 G+24 FE BROADCOM BCM 5633 SERDES Prestera EX 241 $600 24 GE BCM 5345 0. 13 um BCM 5615 M&A(Allayer, Si. Byte) 2002 2003 58

New Roadmap 59

60

So. C 산업의 성공적인 미래 • Global Trend • 중국 시장의 부상 • Eco-System 측면에서 시스템 산업, Foundry • 제품화 Mind의 기술 개발 - Product Planning • Trend 읽기 – 문화 • 기술 마케팅 – TCP, 시장과 성당 • 교육적인 측면에서 61

Global Trend FTA 로 인한 경제 글로벌화 가속 § § IT 버블 붕괴, 인력구조조정 일본- 통합화, 미주/유럽-공동개발 Alliance 활발 Foundry 산업 § § 12” Fab. 건설비용 25억불, AMD+UMC, STMicro+Philips 합작 추진 Infineon SMIC 투자 차세대표준주도권 § Core Open Standard화 -ARM, MIPS, SH, TI, Star. Core, Trimedia 중화권 급부상 반도체 시장 아시아 비중 증가 (2002년 40%) § 아시아 시장의 重要 성 대두 - “왜 우리가 한국 등 아시아에 Behind 되고 있는가 ? ” “연방정부가 나서야 한다. ” [NGN, Broadband session, 2002] § 중국 2005 275억(세계시장의 14%), 미/일본-비중감소 § 대만 : Foundry-> So. C로 사업확대, Fabless Design House (> 1억$ ) 10개이상 § 중국 : Foundry 본격진출, 대만에서 중국으로 Foundry 위주로 산업 개편 § 미국/한국/일본등 - 위기로 대두 § 62

Network Eco System 네트워크 생태계의 건강의 중요성 – Otherwise, arteriosclerosis § § BCN (Broadband Convergence Network) 성공적 로드맵 서비스의 발전, 표준화, 시스템벤더, So. C 의 구현 TCO 최적화를 위한 Embedded Software 및 Foundry 발전의 연계 So. C 발전을 위한 테스트베드 시스템의 투자 e. Bay AOL Daum Yahoo NHN KT China MCI Comcast Telecom CISCO Nortel Redback Huawei Vitesse AMCC Intel Broadcom Agere PMC Marvell Motorola Bandwidth Service, Culture Creation Revenue Generation, Value Creation, Recovery End-to-End intelligence, TCO Minimization Silicon cycle out-sourcing, End to End Solution, Market Window 63

Fabrication § § Foundry Downturn – outsourcing strategy from foundry China semiconductor market $16 B (02), $31 B(06) [Gartner Dataquest] Big 3 – TSMC, UMC, … > $8. 4 B, 90% total Prospect Foundry sales $25 B (07) [Scottsdale IC insights] TSMC § § $4. 6 B (02), 28% , 40% market share $2. 57 B (10 th, 1 H 03) cf. Intel $12. 21 B (1 st, 1 H 03) 0. 18 um + 0. 18 um > 85 % sales Altera, ATI, Motorola, n. Vidia, Broadcom, TI, Via UMC § § Fab alliance with AMD Infineon $3. 6 B partnership IBM § § § Next serious player 6. 1% market share, became Rank 4 th 3 rd, (02) AMD, n. Vidia, Xilinx, Qualcomm, Broadcom SMIC § § Cost advantage 300 mm wafer at 0. 11 um, investment by Infineon Chartered partnership for < 90 ns Broadcom, Elpida, Fujitsu, Infineon, Toshiba 64

65

중국 선진국의 돈과 새로운 에너지가 모이는 곳 상해 § § § 면적 6000 km^2, 1700만 90. 4 포동개발 시작, 연평균 20% 전체 경제 성장율 12% 황하와 천안문사태 연안도시 상해 중점개발, 효과를 몸통(장강) 통해 사천 성 (꼬리) 등 내륙까지 미치게 한다는 경제발전 전략 개인적인 견해 § § § 미국의 주적으로 부상 10년전 일본에 의한 경제헤게모니로 인한 미국 자존심의 상처는, 10여년후 중국에게서 올 수도 엔지니어로 계속살아 남으려면 중국어를 해야할지도 66

북경지역 텐진 내륙지역 상하이 (포동) 반도체, IT 저장성 홍콩 67

68

Education Demand Creation § § § Story, Contents Digital Life – Emotion, Temptation Culture 에 대한 이해 Feature vs. Benefit § § TCP (Target Customer Profile) - 마켓의 소리듣기 TCO (Total Cost of Ownership) Engineering Education § 안목과 소양 l l § 기술적 능력 l l l § 문화와 역사, 인류학 기술정책, 산업전략, 상품화 능력 Technology Leadership Technical Marketing 기술의 경쟁적 Positioning 제품화 Mind의 기술개발 l Product Planning l 고객을 공동의 개발자로 생각하라 [시장과 성당 (cathedral and bazaar) ] 69

맺음말 - 변화와 위험 인생에서 가장 큰 위험은 위험을 감수하지 않는 것이다. “So, first of all, let me assert my firm belief that the only thing we have to fear is fear itself– nameless, unreasoning, unjustified terror which paralyzes needed efforts to convert retreat into advance. ” Saturday, March 4, 1933, Franklin D. Roosevelt At the time, he was speaking of the crisis that was facing the nation —the “Great Depression. ” Today, we reflect on his word, in light of the crisis that now confronts this nation. Joon. Seo@paion. com 70