ARM Technology Update
Dec 2009
1 CONFIDENTIAL ARM CPU Public Roadmap
Market-proven technology ARMv7-Cortex § x1-4 § 200+ partners Cortex-A9 § 600+ licenses Cortex-A8 ARMv6 x1-4 § 15Bu+shipped ARM11™ MPCore™ x1-4 ARM1176JZ(F)-S™ Cortex-A5 ARM1156T2(F)-S™ Cortex-R4F ™ ARM1136J(F)-S Cortex-R4 ARMv5 ARM926EJ-S™ ARM968E-S™
™ ARM7EJ-S™ ARM946E-S SC300™ ARMv4 Cortex™-M3 ARM922T™ Cortex-M1 SC100™ ARM7TDMI(S)™ Cortex-M0
2 CONFIDENTIAL Mali: 无处不在的图形应用
Wireless Mali GPU Roadmap
20+Mali GPU 150+ devices Licensees from OEM graphics partners
x1-4 Automotive Mali-400 MP
Mali-200 Home Roadmap of increasing Key R&D Activities: performance and functionaClityPU+GPU+Fabric Video postprocessing New APIs: OpenCL Mali-55 OpenWF public licensees of Mali GPU
3 CONFIDENTIAL 3 协同合作: Mali 和开发者
Mali Developer Relations & Tools § Content & applications partnerships § Critical & open tools § Emulation or hardware
Mali Middleware § OEM partnerships § Key use cases § System perspective
Mali GPU + Device Drivers § Silicon partnerships § Simple integration § SoC perspective
4 CONFIDENTIAL 4 Cortex: ARMv7 Architecture (A profile) § Thumb-2: Power Efficient Integer Execution § 30% smaller when starting from ARM code § 30% faster when starting from Thumb code
§ TrustZone: Trusted Secure Environment § Device integrity, Digital Rights Management, Electronic payment,etc § Wide industry support
§ Jazelle-RCT: Run Time Compilation Target § Efficient target for Java, Microsoft .NET MSIL, Perl, Python etc § Optional DBX Java byte code accelleration § Early Adopters include Sun Microsystems, Aplix and Esmertec
§ NEON: Multimedia and Signal Processing Architecture § Significant performance uplift from ARMv6 SIMD § Supports both Integer and Floating Point SIMD § Accelerated software development with compiler, library and standard API’s
5 CONFIDENTIAL 5 Cortex-A8
§ High Performance Applications processor § Superscalar pipeline offers 2,000+ DMIPS § Integrated L2 Cache with configurable size (0K-1MB), ECC § High performance with excellent code density § Thumb-2 hybrid 16-32-bit instruction set § Architecture extensions for CPU and system security § TrustZone™ for Secure transactions and Digital Rights Management (DRM) § Multimedia and Signal Processing Architecture § NEON™ provides over 2x Performance of ARMv6 SIMD § OpenMAX library and API’s for fast software development § Efficient Run Time Compilation Target § Jazelle-RCT: Target for Java. Memory footprint reduced up to 3x § Can also target languages such as Microsoft .NET MSIL, Python
65 LP process 65nm G+ process
PPA Optimized Synthesized Optimized Synthesized
Standard Cells Advantage-CE Advantage-HS Advantage-CE Advantage-HS
Memories Custom Advantage** Custom Advantage** RELEASE STATUS Frequency ( MHz ) 650-700 500 1 GHz+ 820MHz
Area with cache (mm2) ~3.3 3.4 ~3.3 3.4 § In Production (Release r3p0) Cache size 32K/32K 32K/32K 32K/32K 32K/32K § Silicon proven since 2006 Power with cache (mW/MHz) 0.6 0.69 0.5 0.61
** Optimized instances of Advantage RAMS Optimized data scaled from 90G Area includes L1 RAMS, L2 control. Excludes NEON, ETM, L2 RAMS Frequency numbers are unmargined
6 CONFIDENTIAL Process is no longer providing scaling
x faster
Lower power
7 CONFIDENTIAL 7 Delivering higher MHz is Expensive § Process shrink is no longer providing increased MHz
§ Higher frequency processors require more complex designs § Exponentially increasing cost; recent modelling has shown to gain 15% MHz, reduces performance around 30% § Harder to design, build and verify, making solution very costly § The resulting increased pipelining lowers performance and increases power consumtion
§ Increasing voltage to increase MHz exponentially increases power consumption
8 CONFIDENTIAL 8 Performance and Power Scalability Supply Voltage § Voltage scaling affects Overdrive peak performance Nominal 4 CPU Multicore can provide Low Power § scalability over performance and power 3 CPU § Larger single core processors typically provide less scalability 2 CPU Single Larger CPU over both power and Multiple Smaller CPU R e l at i ve P r f o m a n ce performance
1 CPU § Larger processors typically are less power efficient at providing a given performance Relative Power Consumption
9 CONFIDENTIAL 9 Cortex-A9 – Market-Driven Solutions
Next-Generation Devices Cortex-A9 Solution Mobile Handsets Next-generation high-end devices (1500-3000DMIPS) 2-3 core processor with IEM™ technology and adaptive shutdown Connected Mobile Computers 32K Instruction and Data caches, 256-512K shared L2 cache using PL310, partitioned AXI NEON technology-based Media Processing Engine, coherent GPU Mid-range, cost reduction, (900-1500DMIPS) Single core processor with NEON or FPU 16K or 32K instruction and data caches 128-256K L2 cache using PL310, single AMBA AXI bus Feature-rich mass market (600-900DMIPS) Single core processor with FPU 16K instruction and data caches, single AXI Consumer and Auto-infotainment Consumer: user interactions (800-3000DMIPS) 1-4 core processors giving design scalability across family of devices 32K instruction and data caches with 0-512K L2 cache NEON technology for advanced media and DSP processing Advanced bus interface unit for high-speed memory transfers between on-chip 3D engines and network interface MACs AMP configurations using separate CPU for real-time RTOS Networking / Home Gateways Enterprise market (4000-8000DMIPS) 3-4 core performance optimized implementation 32K+64K instruction and data cache 512K-2MB L2 cache, dual 64 bit AMBA AXI interfaces Consumer devices (800-1500DMIPS) 1x or 2x multicore utilizing coherent accelerators 32+32K instruction and data, with 256-512K shared L2 cache NEON or VFP when offering media gateway or services Embedded Embedded media and imaging (800-2000DMIPS) 2x multicore utilizing coherent accelerators 32+32K instruction and data with 256K shared L2 cache FPU for postscript and image manipulation and enhancement Code migration through selective AMP/SMP deployments
10 CONFIDENTIAL 10 Cortex-A9 MPCore Processor Structure
Multicore trace and debug Multicore Scalability ARM Coresight Multcore Debug and Trace Architecture
PTM PTM PTM PTM FPU/NEON FPU/NEON FPU/NEON FPU/NEON I/F I/F I/F I/F
Cortex-A9 CPU Cortex-A9 CPU Cortex-A9 CPU Cortex-A9 CPU
Instruction Data Instruction Data Instruction Data Instruction Data Cache Cache Cache Cache Cache Cache Cache Cache
Snoop Control Unit (SCU) Generalized Accelerator Interrupt Control Coherence Cache-2-Cache Snoop and Distribution Timers Port Transfers Filtering
Advanced Bus Interface Unit
Primary AMBA 3 64bit Interface Optional 2nd I/F with Address Filtering High memory bandwidth
ARM MPCore L2 Cache Controller (PL310) Technology
11 CONFIDENTIAL 11 Offering Design Scalability § Delivering over 8000 peak 4 CPU aggregate DMIPS with 8 Product Design Space 2.50 DMIPS/MHz per CPU 7
6 3 CPU § Higher performance and 5 lower power consumption for 4 2 CPU mobile devices 3 s o f D M I PS ( a g r eg t e) ’ 2
1 0 00 1 CPU § Providing software portability 1
across multiple devices Power Consumption / Silicon Cost
A single processor that provides § Synthesis flexibility to target a single product that can be from 500MHz to over 1GHz applied across multiple markets, sharing a common software in TSMC 65nm processes platform
12 CONFIDENTIAL 12 Introducing the Smartbook
Netbook/Notebook Smartbook § Intel’s netbook name is restrictive and positioned as the non-computer § Consumers’ view is a cheap PC § Limited HW specs by Intel (HDD, RAM, Screen Size) § Primarily web browsing and e-mail P r i ce Use/Days of Standby § Always looking to recharge Hours of Use/Day of St- andby -Day All § ARM embraces “netbook” for low-cost and will drive it into price sensitive markets § Still web browsing and e-mail but… § All-day use model WLAN WWAN WWAN § HD multimedia Hotspot Connect on Always Connected § Extremely Thin and Light WiFi boot/resume Your mail and web and download with the flip of a lid § ARM sees the Smartbook class, which leverages the “smartphone” halo § Still web browsing, e-mail; adds productivity applications § Best in class connectivity --Always on / always available § HD multimedia § Sleek form factors
13 CONFIDENTIAL ARM Netbook Silicon: Differentiation Better Power, Better Cost, Better Choice… Today!
§ Freescale i.MX515 (ARM Cortex-A8) § 2D/3D Graphics, HD 720p Video § 512MB DDR2 § 8.9” Display § Integrated 802.11b/g, Bluetoohth 2.0 + ER § <850g § Estimated BoM $120
§ Six public ARM silicon partners for Netbook today § SoC costs between $10 and $20 § Capable of three 2 hour HD movies, 9+ hours web browsing
14 CONFIDENTIAL Lowering System Costs Netbooks -Intel Atom Emerging Devices -ARM 33mm Po15wmmered®
EDGE DDR2 Proc uC Fabric & L2 3G PeriphCache
57 mm RAM 15Audmmio 15mm DSP GPU RAM LTE USB/ Video MIPI
§ Qualcomm® Snapdragon 8250 Feature list • 1GHz ARM v7A CPU • 600MHz aDSP • Support for Android, Linux® & Windows Mobile® External device • Integrated WWAN(3G), support for Wi-Fi & Bluetooth Additional: External device • Seventh-generation gpsOne® engine Cost Decode only • High-definition video decode & encode (720P) Power • OpenGL ES 2.0 core • High-resolution WXGA display support Area No hardware support • Up to 12-megapixel camera
15 CONFIDENTIAL The ARM® Processor In Phone and Now Computer
4000 ARMv5 ARMv6 Cortex™ Cortex-A9 MPCore Scalable computing points required for segmentation
3000
2000 Palm Pre World’s first Cortex-A8 phone
Cortex-A8 Cortex-A9 1000 FOMA902 first phones to ship with ARM11
P e r f or m a n ce D M I PS family processor
ARM11 500 ARM11
ARM9 Feature phones reach 250 MHz on ARM9 family processor
2006 2007 2008 2009 2010 Worst case conditions
16 CONFIDENTIAL Mobile Computing Solutions/Trends HD V i d eo Li n ux A nd r oid E m b e d ed W ind ow G ra p hi cs 3D Ra dios
Company Product Core Speed
ARM11 NVIDIA Tegra™ 650 800 MHz X X X X X MPCore™
Samsung 6410 ARM11 667 MHz X SD X X X X
Samsung S5PC100 Cortex-A8 833MHz X X X X X
V7 Architecture Qualcomm SnapDragon™ 1 GHz X X X X X X License
Freescale iMX515 Cortex-A8 1 GHz X X X X X
TI OMAP™ 3 Cortex-A8 1 GHz X X X X X X 2 cores, TI OMAP™4 Cortex-A9 X X X X X X 1GHz ea PXA3## V5 Architecture Marvell (contact for all 803 MHz X SD X X X X License the products)
MID/Hybrid Emerging Devices V7 – Cortex-A8, minimum 256MB, start of V7 – Cortex-A8 and first Cortex-A9, minimum 512MB 2009 integrated radios and GPS
V7 – Cortex-A8 and Cortex-A9, minimum 512MB, V7 – Mix of single and multi-core, multi GB memory, highly integrated and cost reduced solutions … 2010 start of integration for cost reduction radios, GPS, accelerometers
17 CONFIDENTIAL 18 CONFIDENTIAL CoreMarks: Absolute Performance
ØØCCorortteexx--AA88 proprovviidedess ccoommparabparabllee perperfforormmananccee ttoo IInntteell AtAtomom ØØFFiirrsstt GGeneraenerattiionon CCorortteexx--AA99 dedevviicceess wwillill proprovviidede ccoommparabparabllee perperffoorrmmaannccee ttoo duaduall--ccoreore AAttoomsms ØØNNeexxtt rerevviissiionon ooff CCorortteexx--AA99 wwillill proprovviidede dedesskkttopop lleevveell perperffoorrmmanancece
EEMBC CoreMark at device speed
19 CONFIDENTIAL ARM Computing Ecosystem
ARM Tools, Fabric IP, Models, High-quality Physical IP platforms
Full desktop-class support is enabled Browser has become an “application platform” ARM Cortex-A8/Cortex-A9 + Mali deliver PC class performance. Web is SW not HW dependent Power Consumption is critical HSDPA and WiFi a reality today
OS Enabling Middleware Developers
20 CONFIDENTIAL Computing OS Enablement
Product Operating Systems
“Smartphone” Symbian Windows Mobile Linux -Various Windows Embedded “MID” Linux – Linux.OnARM.com PMP, PND, Internet Tablet Android
Linux (various flavors) Netbooks Ubuntu – Canonical Big Windows -TBD
21 CONFIDENTIAL Web 2.0 Eco-system Updates
Web 2.0 Programs
Mozilla FF3 ARM optimizations feeding into Mozilla project. FF3 & Fennec contains ARM optimizations for JavaScript, FF3.1 will include JIT support for parity with X86. Fennec UI for mobile browsing to complete. Example code on linux.onarm.com Adobe Flash Player10 Joint program with Adobe to optimize Flash Player 10 and AIR and AIR for the ARM architecture. Large improvements for Actionscript and Rendering targeting ARMv6/v7 and OpenGLES 2.0, Adobe Release~ Q4’09
Java Standard Edition Java Standard Edition v6.10 (ARM) available from Sun for ARM Q1 2009 for ARM Linux 2.6.22 and above. Full desktop Java, includes ARM JIT.
Open JDK OpenJDK available with Debian Lenny (5.0) is in Ubuntu 9.04. Contributing VM optimizations for Ubuntu 9.04 release, then adding JIT support
22 CONFIDENTIAL ADOBE FLASH PLAYER 10 SUPPORT ON ARM
23 23 CONFIDENTIAL Adobe Flash Player Technology § The worlds most popular web Rich Content delivery vehicle
70% of all video on Supported by all Web is in Flash leading web browsers 85% of the worlds most popular 99% of all Internet desktop websites use Flash content computers have Flash Player § 1 billion mobile phones will run Flash Player by end of 2009
24 CONFIDENTIAL Adobe Flash Player Technology § Key Flash Player features and supporting technologies
Vector & raster graphics Bi-directional audio & Execution of ActionScript rendering via software video streaming via programs via Tamarin and hardware engines Audio and Video codecs Central Virtual Machine
25 CONFIDENTIAL Comparing Flash and Flash 10 Player
Flash Flash 10 Player
Hardware Processor Class ARM9 ARM11 (>500 Mhz) Requirements Requires VFP No Yes
Basic RAM Footprint 2Mb 40Mb
Compatibility H.264 Video Yes Yes
ActionScript V2.0 V3.0
Plays > 85% of existing Yes Yes § Flash available for web content worldwide mainstream ARM SoCs Application User Interface Yes No Areas today Standalone Applications No Yes § Supports >85% existing Flash content on Web Web content Yes Yes § Provides full H.264 video Availability Release Now End of Q4 § Flash 10 Player schedule in Royalty Free No Yes synch with volume Target Market Feature Phones Yes No availability of Cortex-A class Markets based devices Smart Phones No Yes
Consumer Electronics Yes Yes
26 CONFIDENTIAL Adobe Flash10 on ARM § First Major Milestone for ARM/Adobe Flash10 project § VFPJIT Optimizations for ActionScript 3.0 (2.75 X over interpreter) § Software rendering optimizations bring ARM per/MHz equal to Atom § Flash10 internal demo on ARM1176 and Cortex-A8 platforms § Adobe Argo release schedule now agreed § Open GLES 2.0 is committed for Argo branch §Flash 10.1 public Beta on the web Argo Schedule 2009 2010 §Flash 10.1 release to OEM/ODM’s
Flash 10.1 Argo Adobe Adobe QA and §Flash 10.1 Engineering on documentation released on Web Flash 10.1
Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb
§ Public Beta support for WinMobile and Limited Linux Combinations
27 CONFIDENTIAL Mozilla Browser and nanoJIT Contribution
§ Mozilla Browser § Code engine (NSPR) optimized by 5-20% § Next steps include pushing Neon optimisations in pixman library (Pixel Manipulation Library, part of Cairo project 1.6 and X.org) 1.4 Expect factor 4 to 8 performance gain § 1.2 on unitary tests 1 § Adobe Flash VM (a.k.a. Tamarin) 0.8 before Reached interpreter/JIT ratio of 2.75X for § 0.6 after ARM vs 2.80X for Atom (clock for clock) 0.4 § Improved performance of 48% (geomean) on Sunspider benchmark on Cortex-A8 0.2 (collection of micro kernel benchmarks) 0 Sunspider § Same patches are being integrated into the Mozilla browser code base
28 CONFIDENTIAL -Open Source Java § Optimizing to deliver adequate performance of Java in 2009 § Small applets in a browser (e.g. Facebook Photo upload) § Large scale applications on Linux Desktop (e.g. ThinkFree Office) § ARM optimized Virtual Machine (VM) ready for Ubuntu 9.04 § Recoding “Zero Interpreter” yielded 2X performance for Java Applets § Initial support in Ubuntu 9.04 Release § Optimizing Free version of Sun Template interpreter § Firefox plug-in for small Java applets § Estimate performance improves from 2X to 2.75X for Java Applets § Targeting Beta support for Ubuntu 9.10 § JIT support targeting Cacao JIT for full Java apps § CacaoJIT is about 70% performance of Sun JIT § Not suitable for applets due to memory consumption § Targeted at Large scale applications § Targeting Beta support for Ubuntu 9.10
*Benchmarks = Embedded Caffeine Mar ( which remains the best VMtest)
29 CONFIDENTIAL Linux and Communities on ARM
Linux kernel GNU Tools
30 CONFIDENTIAL Android
http://source.android.com/ http://developer.android.com/
31 CONFIDENTIAL Supporting Two Development Models 1. Embedded Platforms 2.Open Platforms § Incl MIDs, PDA etc. § Incl Netbooks, Notebooks, § Resource constrained Servers… env § Native code development cycle § Relatively closed products § High-spec platform (memory, MHz) § Limited storage § Products easy to reconfigure § Build and test process and expand is relatively complex § Access to disk storage § Cross-development model § Build on PC/x86 and test on target platform or simulator
32 CONFIDENTIAL ARM Linux Internet S/W Platform
§ Leverage MID andNokia Maemo™ project § Use Debian Linux as foundation § Integrate GNOME components and Mozilla web browser § Use reference platforms to validate integration of S/W components § Support ARMv6 and ARMv7 § Facilitate development with group of ARM reference platforms § Website: http://linux.onarm.com
33 CONFIDENTIAL http://linux.onarm.com - Development Model
§ Emulate native development environment using sandbox environment on host PC (x86) § Use tested and reliable configuration scripts § Facilitate importing x86 applications to build them for ARM § Tests can be run directly on emulation platform (Qemu/Fast Models) § Build process should work as with native environment (find correct libraries etc.) § Software can be installed to create a root image § Scratchbox is used to provide CPU transparency enabling multipletargets § Qemu is used to simulate ARM platform § Use GIT repositories to track mainline projects § Provide reference set of projects tracking respective upstream projects § Demonstrate platform inter-operability § Software running on TI, Samsung etc… § Optimized codecs/accelerators may be used on individual platforms but also enable software solution on ARM CPU
34 CONFIDENTIAL Ubuntu Distribution on ARM § Ubuntu 9.04 Jaunty version released § http://www.ubuntu.com/arm § Use i.MX51 as reference platform for Jaunty release cycle § Use the same process and release cycle as for x86 § ‘Karmic’ October’09 upcoming release § Built with ARMv6 ISA + VFP options § Include support for additional HW § Grow ARM Linux ecosystem § Build on Debian “armel” distribution § Provide supported distributions with improved architectural support
35 CONFIDENTIAL Ubuntu -Full Linux Desktop Support
§ Ubuntu is largely based on Debian and GNOME § ARM is now a fully supported architecture with Debian and Ubuntu § Reach a wider developer community to develop & test applicationson ARM
§ Ubuntu Desktop Edition Features include § Gnome based desktop UI, internationalisation and accessibility options § Open Office suite including support for word processing, presentations and spreadsheets § Support for web browsers (incl. Mozilla, Webkit etc.) § Large pool of applications (instant messaging etc) § Primary targets are notebooks and desktops
36 CONFIDENTIAL Projects under Development
§ Open Source Java runtime § ARM contributed to improving OpenJDK performance on ARM platforms § Contributed to improving the virtual machine loop § Close to performance with SUN JRE with JIT § Mozilla and browser support § Contributed some architecture patches to Mozilla code base § Firefox 3.5 already include significant performance improvements for ARM § Enabling Flash 10 support on ARM § Contributed ARM support into Tamarin nanoJIT § Developments also benefiting Mozilla § Enabling use of multimedia and graphic acceleration on ARM SoCs
37 CONFIDENTIAL RVDS 4.0 -Linux support § Simplified building of Linux applications with the ARM Compiler § Support added for GCC command-line options § Upgraded supports for GCC language extensions to GCC-4.2 § Support added for Linux prelink utility § Support for GNU Label variables (goto *table[n]) § Used RVDS to build Mozilla Mobile browser
OMAP 3530 – Cortex-A8
38 CONFIDENTIAL Using ARMCC with Scratchbox § Provide RVDS users with an application to generate an ARMCC tool chain plug-in for Scratchbox. § Any build of the ARMCC tool chain can be used for the plug-in. § A scratchbox-toolchain-armcc-4.0-… tarball is created that can then be installed in Scratchbox. § The plug-in generator can be downloaded from http://connect.arm.com by RVDS users.
§ ARMCC can be used transparently in Scratchbox § Enabled as we added support for GCC command line flags and major extensions. § Enable building with ARMCC of projects which were initially created for GCC. http://www.scratchbox.org/
39 CONFIDENTIAL ARM GNU Tools § ARM is developing patches and contributing direct into GCC mainline § Working with CodeSourcery (www.codesourcery.com) to develop and contribute changes upstream and build stable compiler releases § Track main GCC trunk for ARM architecture support § Track ARM architecture roll-out § Currently support architectures up to ARMv7 § Added support for Cortex-A9 § Latest release is 2009Q1 CS release § Optimisation for some popular ARM µ-architectures § Adding further optimizations for ARMv7 targets § Provide long-term binary compatible releases of the GNU tools for ARM
40 CONFIDENTIAL A Website You Should Read
http://elinux. org/Android_ on_OMAP
41 CONFIDENTIAL Thank You
42 CONFIDENTIAL