2024龍蜥大會英特爾分論壇：英特爾?至強? 6 處理器介紹-趙國棟.pdf

《2024龍蜥大會英特爾分論壇：英特爾®至強® 6 處理器介紹-趙國棟.pdf》由會員分享，可在線閱讀，更多相關《2024龍蜥大會英特爾分論壇：英特爾®至強® 6 處理器介紹-趙國棟.pdf（31頁珍藏版）》請在三個皮匠報告上搜索。

1、英特爾第六代至強可擴展處理器介紹英特爾數據中心及人工智能事業部 首席工程師趙國棟英特爾至強6 處理器介紹趙國棟英特爾數據中心及人工智能事業部 首席工程師至強6概述性能核能效核負載考量0101至強6新特性全新的Die Package設計處理器新特性平臺新特性0202客戶收益性能收益能效收益工作負載示例0303至強6概述性能核能效核負載考量P P-core core SeriesSeries核心性能核心密度計算密集型和AI 工作負載常規用途工作負載高密度和橫向擴展工作負載Optimized for Performance針對每核心性能進優化設計Optimized for Efficiency針對每

2、瓦性能優化設計通用平臺及基礎架構共享軟件堆棧E E-core core SeriesSeries滿足各種性能和效率要求的處理器IntelXeon6 專為滿足市場需求而設計MEMORY SUBSYSTEMSCALAR ENGINEVECTOR ENGINEMATRIX ENGINEOUT OF ORDER SCHEDULEROP SCHEDULERFRONT ENDPM CONTROLLERP P-corecoreE E-corecoreMEMORY SUBSYSTEMSCALARENGINEVECTOR ENGINEOUT OF ORDER ENGINEFRONT END微架構優化性能核（P-

3、Core）能效核（E-Core）E E-corecoreSeriesSeriesP P-corecoreSeriesSeries滿足各種業務負載需求CAECRM,ERPIn-Memory AnalyticsGenerative AIInferenceHCIStorageMedia&GamingModeling&SimulationBig DataDeep LearningMachine LearningVirtualizationCDNVideoEdge AnalyticsConsumer Digital ServicesCloud-NativeApplication DevOpsScale-

4、Out AnalyticsUnstructuredDatabasesCloud-Native CDNNetwork MicroservicesVirtual Protection RelayStorage5G CoreWeb&MicroservicesHPCDatabase&AnalyticsInfrastructure&StorageNetworkingEdgeAIWorkload8LAUNCHED June 2024LAUNCHED June 2024Q12025Q12025Intel Xeon 6900EIntel Xeon 6700PIntel Xeon 6500PIntel Xeon

5、 6 SoCIntel Xeon 6300PIntel Xeon 6700EQ32024Q32024Intel Xeon 6900PIntelXeon6 處理器路標至強6概述全新的die package設計處理器新特性平臺新特性IntelIntel XeonXeon 6 6 微架構UPIUPIUPIUPIMemoryMemoryMemoryMemoryPCIePCIeCore and Cache ArrayCore and Cache ArrayAcceleratorsAcceleratorsAcceleratorsAcceleratorsUPIPCIeCore and Cache Array

6、&FabricAcceleratorsI/O FabricUPIAcceleratorsUPIPCIeAcceleratorsI/O FabricUPIAcceleratorsMemoryMemoryIntelXeon6 芯片封裝演進架構的演變第六代至強可擴展處理器6700 E-core series第五代至強可擴展處理器Two-Tile ArchitectureModuleModule-die Fabricdie Fabric支持靈活的模塊組合構建，為客戶提供廣泛的計算能力選擇IntelXeon6 架構模塊化 SoC 架構的可擴展性和靈活性MultiMulti-die Architectur

7、edie ArchitectureI/O die 包括 UPI,PCIe,CXL 和 IntelAccelerator Engines計算 die 包括 cores,cache 和 memory controllersEmbedded MultiEmbedded Multi-die Interconnect Bridgedie Interconnect Bridge 實現封裝內（In-package）高密度模塊互聯 高帶寬、低功耗和低延遲Last level cache 所有核心共享L3緩存，可以劃分為 per-die sub-numa clustersEmiB technology 將高速F

8、abric擴展到各個模塊Modularity and flexible routing 模塊化和靈活的布線允許按芯片定義行和列Fabric distributes 在多個列之間分配 IO 流量以緩解擁塞Monolithic Mesh 支持Socket內代理之間的直接訪問Global infrastructure 全局模塊化和分層架構Cores,CH A,LLC&M esh Fabri cD D R5/M CRM em oryD D R5/M CRM em ory基于Intel 3 制程 更高的性能和能效模塊化計算芯片架構通用的 I/O StacksUPI,PCIe,CXL and Intel

9、Accelerator Engines新增功能Full CXL support,extends Resource Director Technology(RDT),secure interconnect增強的 I/O 性能UPI 24GT/s w/6-links,UPI affinity,distributes traffic across all mesh columnsI/O FabricUIOAccelUIOIOIOUIOAccelUIOUPICXLPCIeCXLPCIeDSAIAAQATDLBIntel Data Streaming Accelerator(Intel DSA)Inte

10、l In-Memory Analytics Accelerator(Intel IAA)Intel QuickAssist Technology(Intel QAT)Intel Dynamic Load Balancer(Intel DLB)基于Intel 7 進程 通用 I/O 模塊模塊化 I/O 芯片架構Intel Xeon 6700Intel Xeon 6700-seriesseriesIntel Xeon 6900Intel Xeon 6900-seriesseriesP-core and E-core SKU selectionsIncreased core countsIncrea

11、sed memory bandwidth with DDR5 Increased inter-socket bandwidth with UPI 2.0Compute Express Link(CXL)2.0Increased I/O bandwidth on PCIe 5.0Increased shared LLCIntel Accelerator EnginesHW-enhanced securityCommon OS and firmwareMultiplexed Rank DIMM(MRDIMM)Increased Intel VMD domains共享底層平臺 提供靈活的硬件配置選擇

12、英特爾至強6 處理器|6700 和 6900 平臺系列Category Category Software Stack ComponentSoftware Stack ComponentEfficientEfficient-corecore(E-core)PerformancePerformance-corecore(P-core)Instruction Set and ExtensionsBase x86 ISAxxIntel Advanced Vector Extensions 2(Intel AVX2)xxIntel Advanced Vector Extensions 512(Inte

13、l AVX-512)xIntel Advanced Matrix Extensions(Intel AMX)xOS and HypervisorLinux kernel and commercial LinuxxxWindowsxxVMware ESXi xxApplications and Libraries Database mon libraries(ex.ZStd)xxNetwork&media mon libraries(ex.DPDK)xxGeneral compute&storage incl.libraries(ex.SPDK)xx使用通用軟件堆棧提高開發效率和易用性簡化開發和

14、部署Intel Software Guard Extensions(Intel SGX)Common OS and firmwareIncreased shared LLCIntel Accelerator EnginesIntel Trust Domain Extensions(Intel TDX)6900 SeriesUp to 288 Efficient-coresUp to 128 Performance-cores6700 SeriesUp to 144 Efficient-coresUp to 86 Performance-cores 1S/2S and4S/8S(P-core o

15、nly)support4UPI 2.0 links,up to 24 GT/sUp to 350Wper CPU8channel memory Up to 6400 MT/sDDR5 memory8000 MT/s MCR DIMM memory(P-core)Up to 88lanes PCIe 5.0/CXL 2.0(up to 136 lanes for 1S designs)1S/2S supportUp to 500Wper CPU12channel memory Up to 6400 MT/sDDR5 memory8800 MT/s MCR DIMM memory(P-core)U

16、p to 96lanes PCIe 5.0/CXL 2.0 6UPI 2.0 links,up to 24 GT/s靈活可擴展的產品配置英特爾至強6 處理器|6700 和 6900 平臺產品Socket SupportUPI LinksPCIe/CXLMem ChannelsMax TDP6900 Series6700 SeriesDDR5PCIe5UPI 2.0Up to2.3xhigher memory bandwidth(w/MCR DIMM memory in P-core)vs.5th Gen Intel Xeon processorsUp to1.2xincreased I/O B

17、andwidth vs.5th Gen Intel Xeon processorsUp to1.8xincreased inter-socket bandwidth vs.5th Gen Intel Xeon processorsUp to1.4xhigher memory bandwidth(w/MCR DIMM memory in P-core)vs.5th Gen Intel Xeon processorsUp to1.1xincreased I/O Bandwidthvs.5th Gen Intel Xeon processorsUp to1.2xincreased inter-soc

18、ket bandwidth vs.5th Gen Intel Xeon processorsCXL 2.0Type 1,Type 2,and Type 3基于IntelXeon6 平臺性能增強Seamless Firmware UpdateSeamless Firmware UpdateIncreasingg server uptime,reduces user disruptions while deploying server platform updates and mitigating potential security vulnerabilities in a few second

19、sReliability Availability Reliability Availability Serviceability(RAS)Serviceability(RAS)Maximizing system uptime by using features for error detection and correction,runtime repair and recover,resulting in an improved total cost of ownership.Intel Platform Monitoring Intel Platform Monitoring Techn

20、ology(Intel PMT)Technology(Intel PMT)Extensive,flexible,industry-leading platform telemetry for server fleet management and facility orchestrationIntel Flat Memory ModeIntel Flat Memory ModeUsing logic in the CPU to keep frequently used data in DRAM where performance is best and swapping out less fr

21、equently used data to CXL memoryActive Idle Power ModeActive Idle Power ModeLowering Uncore frequency in low activity scenarios where a minimal utilization point is set and thresholds to minimize impact to workload performance.Intel Accelerator EnginesIntel Accelerator EnginesIntel Dynamic Load Bala

22、ncer(Intel DLB)Intel In-Memory Analytics Accelerator(Intel IAA)Intel QuickAssist Technology(Intel QAT)Intel Data Streaming Accelerators(Intel DSA)Optimized Power ModeOptimized Power ModeReducing power consumption with minimum performance trade-off to reduce operational carbon footprintCompute Effici

23、encyOperational EfficiencySimplifying power management by providing an architectural MMIO-based solution designed to PCIe standards which affords flexibility,expandability,software reuse from generation to generationTopology Aware Power Management Interface Topology Aware Power Management Interface(

24、TPMI)(TPMI)基于IntelXeon6 平臺特性增強App IsolationApp IsolationIntel SGXIntel SGXVM IsolationVM IsolationIntel TDXIntel TDXTrust BoundaryCloud Stack&AdminsBIOS&FirmwareHypervisorVM AdminGuest OSEnclaveApplicationsConfidential DataCloud Stack&AdminsBIOS&FirmwareHypervisorVM AdminGuest OSApplicationsConfiden

25、tial DataTrust Boundary信任邊界：具有訪問機密數據潛力的元素。信任邊界之外的元素被阻止訪問內部的元素.Enclave sizes up to 512 GB per processor Enhanced Row hammer protection with Cryptographic Memory Integrity option&AEX-Notify ISA extension Intel Xeon 6 introduces AES-256 encryption(quantum-resistant)基于硬件的數據機密計算安全保護激活敏感、受監管或主權數據資產英特爾機密計算

26、產品組合Intel Software Guard Extensions(Intel SGX)最小信任邊界-機密數據訪問僅限于已證明的應用程序代碼 Generally available starting with 5th Gen Intel Xeon Scalable processors Hardware support for Live Migration of Trust Domains Intel Xeon 6 introduces AES-256 encryption(quantum-resistant)and support for up to 2048 encryption ke

27、ys for trust domains with Intel TDXIntel Trust Domain Extensions(Intel TDX)最小信任邊界-機密數據訪問僅限于已證明的應用程序代碼客戶收益性能收益能效收益工作負載示例See 7W4,7N2,7T2,7D1.9G10,9H10 at Xeon 6.Results may varyData ServicesWeb&MicroservicesAIUp to 2xHigher performance on AI inferenceHigherserver-side perf/watt java throughput w/SLAHP

28、CHigher performance on OpenFOAMworkloadsUp to2.3xUp to 2.6xNetworkingUp to3.4xHigher performance/watt for next gen firewallMediaUp to 2.6xHigher performance/watt for media transcodeUp to 2.7xHigher performance/watt for MySQL OLTP workloadsE E-corecoreE E-corecoreE E-corecoreE E-corecoreP P-corecoreG

29、en Over Gen5 Year RefreshGeneral ComputeUp to2xAverage Higher performance for general compute P P-corecoreP P-corecoreIntelXeon6 出色的性能和效率3to 1機柜整合(2nd Gen Intel Xeon to Intel Xeon 6 with E-cores)提高機柜利用率提升效率和總擁有成本每個Socket 多達144個核心數據服務、網絡、媒體和微服務高核心密度，優秀的橫向擴展能力優于競品1.3倍的每瓦性能高能效（perf/watt）為大規模云scale out工

30、作負載提供獨特的優勢See 7T1,7W210 at Xeon 6.Results may varyIntelXeon6700 E 核處理器利用云原生敏捷性優化和擴展您的基礎設施01002003004005006007008000%20%40%60%80%100%2S Socket Power(W)Lower is betterServer UtilizationIntel Xeon 8592+(64C,350W)Intel Xeon 6780E(144C,330W)*Out-of-Box mode:Assumes default energy-efficient BIOS and OS se

31、ttings.Socket power is power consumed by CPUs280WSee 7T3 at Xeon 6.Results may varyE 核的 IntelXeon6 能效表現（基于負載利用率）基于能效核的 Intel Xeon 6服務器的功耗收益Intel Xeon 6 with EIntel Xeon 6 with E-core vs.5th Gen Intel Xeoncore vs.5th Gen Intel Xeon在具有 E 核的 IntelXeon 6 上，功率隨負載線性增加在最佳位置運行時可節省高達 280W 的功率 40-60%的服務器利用率In

32、tel Xeon 6780E 與Intel Xeon Platinum 8592+相比性能提高 18%使用默認的開箱即用設置降低數據中心的電力和冷卻成本1.031.251.110.940.951.361.421.261.291.341.551.631.201.491.461.521.351.660.00.20.40.60.81.01.21.41.61.82S Intel Xeon 6780E with E-cores(144 cores)vs.2S 5th Gen Intel Xeon Platinum 8592+(64 cores)Normalized to 5th Gen Intel Xe

33、on-SP(Higher is better)Normalized to 8592+Socket power is power consumed by CPUsSee 7G2,7D2,7W5,7W2 and backup at Xeon 6.Results may vary基于E-core 的 IntelXeon6700 vs.第 5 代 IntelXeon處理器典型工作負載 Gen To Gen 收益MediaNetworkWebDatabaseGeneral ComputePerformancePerformance/Watt3.253.663.703.182.964.074.224.22

34、4.175.832.292.592.782.362.602.552.642.642.663.460.01.02.03.04.05.06.07.0Intel Xeon 6 E-core performance&efficiency vs.2nd Gen Intel Xeon processorsNormalized to 2nd Gen Intel Xeon-SP(Higher is better)PerformancePerformance/WattMediaIntel Xeon 6780E(144c)vs.Intel Xeon 8280(28c)NetworkIntel Xeon 6780E

35、(144c)vs.Intel Xeon 6252N(24c)WebIntel Xeon 6780E(144c)vs.Intel Xeon 8280(28c)DatabaseIntel Xeon 6780E(144c)vs.Intel Xeon 8280(28c)General ComputeIntel Xeon 6780E(144c)vs.Intel Xeon 8280(28c)See 7G1,7D1,7W4,7N1,7N2 at Xeon 6.Results may vary基于E-core 的 IntelXeon6700 vs.第 2 代 IntelXeon處理器典型工作負載平臺更新性能和

36、能效收益（5-year refresh）*STREAM Triad workloadSee 7N5 at Xeon 6.Results may vary升級系統使CDN 視頻點播業務獲得24倍及8倍的性能及能效收益基于IntelXeon6 平臺收益示例2.5x more threadsLess power per threadMemoryPCIe Network/Storage Intel Xeon 6780E vs.Intel Xeon 6780E vs.2nd Gen Intel Xeon 82802nd Gen Intel Xeon 82804x NVMe drives4x Ethern

37、et connection speeds4x PCIe Lanes 4x PCIe transfer rate per laneLower PCIe controller latency1.3x more mem channels2.2x increase in DRAM speedUp to 2x Mem bandwidth*CPUVideoVideo-OnOn-DemandDemandXeon 6780E Performance and Perf/Watt vs.2nd Gen Intel Xeon 8280(800 Connections per client,Normalized to

38、 8280)24x24x8x8x0.05.010.015.020.025.030.06780E Perf6780E Perf6780E Perf/Watt6780E Perf/WattInteger ThroughputInteger ThroughputIntel Xeon 8280 vs.Intel Xeon 6780EMedia Transcode(AVC)Media Transcode(AVC)Intel Xeon 8280 vs.Intel Xeon 6780EIntel Xeon 6700E服務器性能提升（每臺）3.6x3.6x4.2x4.2xIntel Xeon 6 700E服務

39、器能效提升（每臺）2.7x2.7x2.6x2.6x機架整合3:13:1每個整合機架4 年可節約能源1,540 MWh1,195 MWh每個整合機架4 年二氧化碳排放量減少650 tons500 tonsIntel Xeon 6700E 系列的1 個 15KW 機架與基于 Intel Xeon 的第二代服務器的 3 個機架的比較See 7T1,7T2 at Xeon 6.Results may varyIntelXeon6 顯著減少了數據中心基礎設施的空間、電力和成本更新舊服務器，獲得顯著的機架級優勢80k MWh1Over 4 yearsFleet energy saved34k mt1Red

40、uced CO2 Emissions29IntelXeon6700E3:13:1Rack Consolidation12nd Gen IntelXeon Processor之前200 機柜66 機柜現在See 7T2 at Xeon 6.Results may vary在更小的空間和更低的功耗下承載相同SLA的普通業務流，以支持新的 AI 項目應對AI 數據中心能耗限制及挑戰SKUSKUCoresCoresMicro Micro ArchitectureArchitectureBaseBase(GHz)(GHz)All CoreAll CoreTurbo(GHz)Turbo(GHz)MaxMa

41、xTurboTurbo(GHz)(GHz)L3 CacheL3 Cache(MB)(MB)TDPTDP(Watts)(Watts)Max Max Scala.Scala.DDR5 MemoryDDR5 MemorySpeedSpeed(1 DPC)(1 DPC)Default Accelerator Default Accelerator DevicesDevicesIntel TDX Intel TDX Keys Keys(Per CPU)(Per CPU)Long Long Life Life AvailableAvailable*UPI UPI Links Links EnabEnab.

42、PCIe5 PCIe5 Express Express Lanes/CLanes/CXLXL6780E144E-core2.23.03.01083302S64002 Intel DSA,2 Intel IAA,2 Intel QAT,2 Intel DLB,6 Intel VMD10244886766E144E-core1.92.72.71082502S64002 Intel DSA,2 Intel IAA,2 Intel QAT,2 Intel DLB,6 Intel VMD10244886756E128E-core1.82.62.6962252S64002 Intel DSA,2 Inte

43、l IAA,2 Intel QAT,2 Intel DLB,6 Intel VMD1024-4886746E112E-core2.02.72.7962502S56002 Intel DSA,2 Intel IAA,2 Intel QAT,2 Intel DLB,6 Intel VMD1024-4886740E96E-core2.43.23.2962502S64002 Intel DSA,2 Intel IAA,4 Intel QAT,4 Intel DLB,6 Intel VMD10244886731E96E-core2.23.13.1962501S56002 Intel DSA,2 Inte

44、l IAA,2 Intel QAT,2 Intel DLB,6 Intel VMD1024-0886710E64E-core2.43.23.2962052S56002 Intel DSA,2 Intel IAA,4 Intel QAT,4 Intel DLB,6 Intel VMD1024488Intel may make changes to specifications and product descriptions at any time,without notice.Please visit or contact your Intel representative to obtain

45、 the latest product specifications.Intel processor numbers are not a measure of performance.Processor numbers differentiate features within each processor family,not across different processor families.All processors support Intel Virtualization Technology(Intel VT-x).*Long Life Availability:7+yearsIntelXeon6 E-core 處理器型號