Impact Of Technology Trends On Computer Architecture

IBM ResearchImpact of Technology Trends onComputer ArchitectureJaime H. MorenoIBM Thomas J. Watson Research CenterYorktown Heights, NYAugust 2008 2008 IBM Corporation

IBM ResearchIBM Research WorldwideOver 3000 Researchers ineight labs around the world2J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchDiversity of Disciplines at IBM Research3Behavioral SciencesChemistryComputer ScienceElectrical EngineeringMaterials SciencesMathematical SciencesPhysicsService Science,Management & EngineeringJ. Moreno / August 2008 2008 IBM Corporation

IBM ResearchIBM Research’s Strategic ThrustsIndustryr SolutionsServr icesServicesSoftf wareSoftware System Architecture System Software Deep Computing Microprocessors and Tools Microprocessor Architecture VLSI Design Design AutomationStorage SystemsSystemsT chnologyTeTechnologyExploratoryr4J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchConversation at a party So, what do you do professionally? I am a Computer Architect (pause . silence .) Oh, I see so you design computer cases.How come most of the time they are so boring, even ugly?5J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchArchitecture vs Chip Architecture6J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchImpact of Technology Trends on Architecture Science, engineering and technology are constantly providing new materials, tools, methods,etc., leading to changes in the architecture of buildings– Constant evolution of capabilities, enabling all sorts of innovation Similarities can also be drawn with other areas– Cars, for example7J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchKnowing the Technologies and their LimitationsOn November 7, 1940, at approximately 11:00 AM,the first Tacoma Narrows suspension bridgecollapsed due to wind-induced vibrations. Situatedon the Tacoma Narrows in Puget Sound, near thecity of Tacoma, Washington, the bridge had onlybeen open for traffic a few months.“Disasters” may also happen in Computer Architecture, albeit no so dramatically, as a consequence ofthe use of the technologies8J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchImpact of Technology Trends on Computer Architecture What is Computer Architecture?– (Valentina already gave us a good perspective) In one simple term: a “contract”– Conceptually, very similar to the architecture of a building A specification for the “builder” computer engineers A specification for the “user” software engineers But there are some important differences– Technology evolves very fast, so a system looses its advantages ina short period imagine if that was also the case for homes– Although systems change rapidly, expectation is that softwaremigrates from system to system “transparently” Compatibility with “performance” scalabilitySoftwareArchitectureImplementation 1Implementation 2Implementation 3Implementation n Because of the technology changes, a successful “computerarchitecture” normally has many different “implementations”9J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchLevels of Computer Architecture Multiple levels in Computer Architecture– Multiple “contracts” throughout a system– System, Network, Multiprocessor, Processor, Cache memory, Pipeline, In some cases, we use the term “microarchitecture” to refer to lower levels– Microarchitecture closely related to “organization”– One “processor architecture” may have multiple “microarchitectures”– Each “microarchitecture” may have multiple implementations Technology trends impact primarily the implementations, but the effects are oftenvisible at the architecture levels10J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchTechnology Trends Example A simplistic view– Early 70s: Memory was small and slow, compiler technology not mature Complex instruction sets (CISC), “semantic gap,” language-specific machines– Early 80s: Memory improving rapidly, advances in compiler technology RISC (Reduced Instruction Set) processors emerge Semantic gap replaced by sequences of instructions RISC vs CISC debated extensively– Early 90s: Transistor density enables more logic in a chip CISC translated into RISC on-the-fly RISC vs CISC debate became pointless - all RISC internally– Early 00s: Transistor density enables even more logic in a chip Processors becoming more CISCy again .? Key observation– Changes in technology can trigger major transitions in Computer Architecture– Recognizing and anticipating the changes leads to breakthroughs11J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchExamples of Technologies in Computing Logic and Main Memory Storage– Vacuum tubes– Magnetic drums, disks, tapes Input/output devices– Magnetic core memory– Semiconductor transistors and memory– Keypads, switches, teletype writers– Integrated circuits: SSI (small), MSI (medium),LSI (large), VLSI (very large)– Punched cards, punched tape– Bipolar, NMOS/PMOS, CMOS– 32x80 characters display– Dot-matrix printers– Graphics displays, High-resolution devices– Many others– etc., etc. Gallium Arsenide, Bubble memories,etc.Every major technology innovation has brought major changes to thearchitecture of computers and it continues to be soTechnology eras12J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchExamples of Technology Eras in Computing Vacuum tube era– Very little logic, very little memory– Very short time to failure Magnetic core memory era– Non-volatile– Larger capacity, more reliable Semiconductor era– Much larger capacity (logic and memory)13J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchTechnologies define Eras Not one but multiple technologies define an era– Processor and memory– Storage– CommunicationsPrimary focus today(Chip architecture)– I/O devices– . Consequently, eras have different attributes Eras can also be recognized by the usage mode of the computers– Batch– Interactive– Web-driven– Real-time driven14J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchRecent Microelectronics Technology Eras14 Bipolar Transistors Era– Provided continued improvement in thecharacteristics of the systems Increased number of transistors in achip and their speed– However, bipolar transistors consume powerconstantlyCMOS Era12Module Heat Flux (W/cm2)– Dominated industry in the early years ofmodern computersBipolar Era1086JunctionTransistor4IntegratedCircuit2 CMOS Transistors Era0– Consume power only when there is switchingactivity1950196019701980199020002010– Initially, slower than bipolar transistors buteventually became faster– Continued trend of improving performance ateach generation, without requiring changes tothe software Historic trend: 2x every 2 years atroughly constant cost15J. Moreno / August 2008 Easy migration for softwareWidespread adoption of computing technology 2008 IBM Corporation

IBM ResearchMajor Trends: Transistor Density and Frequency101.0E 1010 Where did we go “wrong”?9101.0E 09– Explosion of leakage current1 Billion8101.0E 08 50% CAGR10001.0E 0710Power Density (W/cm2)71006101.0E 061 Million5101.0E 051980101.0E 0419851990199520002005201042004 Frequency ExtrapolationActivePower10Passive Power10.1Gate Leakage0.01199420050.0011.0E 0310130.10.01Gate Length (microns)Gate Length (microns)101.0E 02219901619952000J. Moreno / August 200820052010 2008 IBM Corporation

IBM ResearchCMOS Era: Benefits from Technology Scaling are DiminishingTransistor performancescaling to continue, butat a slower ratePower is limitingpractical performanceSingle thread performanceis slowingNo more “free ride” for software 17J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchDidn’t we have this problem in the past ? Bipolar Era symptoms are back but there is no other semiconductortechnology ready to be used this time at least not for a number of years– Nanotubes– Quantum computing– Molecular computingBipolar EraJ. Moreno / August 2008CMOS 50– .18Module Heat Flux (W/cm2)14196019701980199020002010 2008 IBM Corporation

IBM ResearchA New “Technology” Era - The “CMOS Transition” EraCMOS EraCMOS Transition EraSystem PerformanceBipolar EraNanotech EraNew technologiesMulticoreSingle coreHistoric trend19701990200520102020 CMOS scaling alone can no longer provide simultaneous improvements in all productmetrics (performance, cost, power, software compatibility) at the historic rate. New technologies, systems and software evolution will be required to achieveperformance and cost improvements over the next decade. Computer architectures will have to drive the transition19J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchWhere are we today? CMOS density continues roughly at historicrate while operating frequency is basicallyflat, because of power limitations Technologies other than CMOS needed tocompensate the trends– We can have more transistors per chip butcannot have them run faster– Semiconductor materials and structures,lithography, design layout– If they run faster, we cannot use all of them atthe same time– Operation at ultra-low voltage, advancedpower management Single-thread performance improving slowly Transitioning to multicores, accelerators,specialized systems– Software changes required– Leverage point is moving “up the systemstack”– Cooling– 3D integration– New memory technologies, denser and lowerpower, non-volatile– On-chip optical communications (photonics)– Heterogeneous and specialized engines andsystems20J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchWhy 3D Integration? Multi-core, multi-threaded and multi-image(virtualization) chips are stressing memorybandwidth and on-chip memory capacity Advances in technology providing muchhigher I/O densities and bandwidth betweenchips Mix different types of chips in one package Design challenges: thermal, package I/O, etc. Impact to architecture (example)– No longer constrained by 2D organization– What if we had Gigabytes of cache memoryon chip, and could transfer several Kilobytesper cycle across cache levels?21J. Moreno / August 2008 2008 IBM Corporation

IBM Research3D Integration Restructuring the architecture of the chip / nodeImprovedthermalconductivitythroughstackPhase change memorySilicon interposer22J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchNew Memory Technologies Impact to architecture (example) Phase-change Memory (PCM) attributes:– Fast: Tens of nano-seconds– What if we had Terabytes of mainmemory, but the number of writeoperations are limited?– Low-Power: Non-volatile storage– Dense: Multi-bit, projected to becomparable to disksDRAMPCMNANDFLASHENTERPRISEDISKRead access time (us)0.050.1205000Random Write Access (us)0.050.115005000Device Capacity (GB)0.514032500 - 2500Device Bandwidth (MB/s)1000100020150Endurance1015109 - 10121051012Device Power (W)0.20.10.1 - 0.210-20Legend:23J. Moreno / August 2008Best Barrier 2008 IBM Corporation

IBM ResearchOptical Interconnect Technology Impact to architecture (example)– What if we could transfer data within a system 100 times faster?Carrier24J. Moreno / August 2008 2008 IBM Corporation

IBM ResearchHeterogeneous Systems Emergence of heterogeneous components in processor and systems– Recently announced “Roadrunner” supercomputer has Cell and x86 chips Fastest and most power efficient computer in the world Impact to architecture (example)– What is the architecture of a system with increased heterogeneity (hardware and software)?MulticoreHomogeneousStandardized Interconnect25J. Moreno / August 2008 2008 IBM Corporation