Open Server Summit 2016 : Linley Group Slides
- 1. © 2016 The Linley Group March 15, 2016 ARM Server Processor Opportunity Linley Gwennap, Principal Analyst April 14, 2016
- 2. © 2016 The Linley Group March 15, 2016 About Linley Gwennap • Founder, principal analyst, The Linley Group • Leading vendor of technical reports on processor, mobile, and communications semiconductors and IP • Editor-in-chief of Microprocessor Report • Author of recent articles on ARM, Broadcom, Intel, Marvell, MediaTek, NetSpeed, Qualcomm, et al • Author of “A Guide to Mobile Processors” and “Mobile Semi Market Share Forecast” • Former CPU designer at Hewlett-Packard 2
- 3. © 2016 The Linley Group March 15, 2016 Agenda • Server market opportunity • Three ARM waves for servers 3
- 4. © 2016 The Linley Group March 15, 2016 Clouds Open Arms to New Architectures • Cloud service providers reaching out for non x86 processors • Intel has >97% server processor market share • CSPs want a multi-vendor environment • Not possible for one vendor to cover all application spaces • Intel offer high memory bandwidth only with high CPU performance • Many applications benefit from hardware accelerators • Intel can address these problems with custom products, but only for large customers 4 One Solution Does Not Fit All Applications
- 5. © 2016 The Linley Group March 15, 2016 New Cloud Technologies Change Processors • Microservices break down traditional scale-up applications • Better for smaller CPU cores • Hyperconvergence merges compute and networking in one box • Networking is well suited to smaller CPU cores, hardware acceleration • Reduced demand for multi-socket systems • SoC processors reduce size and power without coherency overhead • New workloads change balance between CPU, memory, I/O • In-memory databases need large memory but moderate compute • Cold storage needs high bandwidth but little compute 5
- 6. © 2016 The Linley Group March 15, 2016 Opportunities for ARM in Servers • Market needs 64-bit processors—ARMv8 now available • Xeon E5 defines mainstream server market • Little to no volume in microservers • Xeon E7 is too expensive for all but a few applications • ARMv8 software ecosystem continues to build • Commercial support for open-source stack (Linux, etc) • CSPs porting their in-house software using GCC • APIs available for networking functions (SDN and NFV) 6
- 7. © 2016 The Linley Group March 15, 2016 First ARMv8 Wave (2014-15) • AppliedMicro X-Gene 1 entered production in 2Q14 • First ARMv8 server processor • First merchant ARMv8 processor product of any kind • 8-core X-Gene 1 used 40nm technology to accelerate time to market • AMD A1100 (“Seattle”) ready for production in 2015 • 28nm processor using 8 stock Cortex-A57 CPUs • First wave products had low performance, targeted microservers • Intel offered 8-core Atom (“Silvermont”) with better perf/watt • ARMv8 server software infrastructure wasn’t ready for deployment 7
- 8. © 2016 The Linley Group March 15, 2016 Second ARMv8 Wave (late 2015-2016) • AppliedMicro X-Gene 2 entered production in 2H15 • Move to 28nm improves perf/watt, but not significantly • Quad DRAM channels offer more memory bandwidth, capacity than Atom • Cavium ThunderX entered production in 2H15 • Massive 48-core design matches per-socket throughput of low-end Xeon E5 • Low per-core performance enhanced by hardware accelerators • Second wave offers advantages over Intel for specific workloads • X-Gene 2 good for scale out memory-intensive applications (e.g. in-memory DB) • ThunderX good for embedded applications, esp. storage and networking 8
- 9. © 2016 The Linley Group March 15, 2016 Third ARMv8 Wave (2017-2018) • AppliedMicro X-Gene 3 targets production in 2H17 • 32 cores, 3.0GHz processor will challenge Xeon E5 • Qualcomm likely to reach production in 2017 • No product details yet, but we expect 32 cores using smartphone CPU • Cavium moving ThunderX to FinFET, production likely in 2017 • Broadcom Vulcan on hold after Avago deal • Third wave targets mainstream Xeon E5 customers • Uses 16/14nm FinFET technology on a par with Intel 14nm “Skylake” • Brawnier cores in X-Gene 3 compete for Web-scale computing 9
- 10. © 2016 The Linley Group March 15, 2016 Requirements for ARM Server Processor Adoption • Match performance of Xeon E5 • Xeon E5 is the most commonly deployed server processor—need to replace • Match per-core and per-socket performance to serve scale-out applications • Match or exceed Xeon E5 power efficiency (perf/watt) • Requires strong microarchitecture and be within one process node of Intel • Excellent cache performance and multicore scaling • Proven reliability, availability, and serviceability (RAS) feature set • Exceed Xeon E5 in one or more areas (e.g., memory, accelerators) 10
- 11. © 2016 The Linley Group March 15, 2016 Conclusions • Server OEMs want more choices in processor suppliers • ARMv8 software infrastructure is maturing in the data center • Early ARM server processors deliver advantages on certain workloads • Hindered by low core count and/or lagging process technology versus Intel • Third wave of ARM server processors offers more choices • High performance, high core counts for scale-out workloads • Hardware accelerators for specialized workloads • Server processors are no longer a single-vendor market 11