Transient Thermal Analysis For M.2 SSD Thermal Throttling: Detailed CFD .

Transient Thermal Analysisfor M.2 SSD Thermal Throttling: Detailed CFD Model vs Network-Based ModelHedan Zhang1, Hainan Wang2, Shay Braha3,Ernold Thompson4, Ning Ye1, Nathan Ai2, CT Kao2, Nir Amir31Western Digital, 951 SanDisk Drive, Milpitas, CA 950322Cadence Design System, 2655 Seely Ave, San Jose, CA 951343Western Digital, Kfar Saba, Israel, 444254Western Digital, Bengaluru, Karnataka 560103Email: Hedan.zhang@wdc.comABSTRACTSolid State Drive (SSD) technology continues to advancetoward smaller footprints with higher bandwidth andadoption of new I/O interfaces in the PC market segment.Power performance requirements are tightening in the designprocess to address specific requirement along with thedevelopment of SSD technology. To meet this aggressiverequirement of performance, one major issue is thermalthrottling. As the NAND and ASIC junction temperaturesapproach their safe operating limits, performance throttlingis triggered and thus power consumption would dropaccordingly. Therefore, robust thermal understanding onsystem level as well as reliable and fast thermal predictionare becoming essential in the process of system thermaldesign to optimize performance in a quick turnaroundmanner.In this paper, we present two different modelingapproaches on the system level to model and simulate M.22280 SSD thermal throttling behavior in a typical laptopworking environment. One approach is to establish a detailedthree dimensional CFD (computational fluid dynamics)model using traditional CFD tools. In this model, themotherboard is enclosed in a case or chassis. Major heatsources of components and packages on the motherboard areconsidered including CPU, GPU, M.2 SSD, DRAM etc.Advanced cooling solutions like heat pipe and blowers arealso modeled. In order to accurately capture thermalbehavior of the SSD, detailed structure and geometry ofNAND, PMIC and ASIC packages are included. Bothnatural and force convection as well as radiation areconsidered in this model. Both steady state and transientsimulation results are presented in this paper. Further, thesimulation results are validated with experimental data topredict thermal throttling behavior. The experiment iscarried out with the SSD running in a laptop andtemperatures of NAND and platform are logged during thetest.In this paper, a second approach to generate accuratethermal models is presented for electronic parts. The thermalmodel of an electronic part is extracted from its detailedgeometry configuration and material properties, so multiplethermal models can form a thermal network for complexsteady-state and transient analyses of a system design. Theextracted thermal model has the following advantages,1. It can accurately predict both static and dynamicthermal behaviors of the electronic parts;2. It can accurately predict the temperature at any probingnode pre-defined in the electronic part;3.It is independent of boundary condition and canaccurately predict the thermal behavior regardless ofthe environment and cooling conditions.With the accurate dynamic thermal models, a largethermal system can be decoupled into multiple domains suchas air flows, chassis, heat sinks, PCB boards, packages, etc.The whole system can be consequently reconstructed as anintegrated model-based network, and thermal simulation canbe performed using fast network simulators. In comparisonto the traditional CFD or FEM tools, the network-basedapproach improves efficiency in both thermal systemconstruction and simulation. This approach is demonstratedthrough thermal simulation of the SSD drive within a laptopenvironment under natural convection in its workingcondition. The simulated system includes packages, M.2PCB, motherboard, heat sink, and chassis.KEY WORDS: network model, laptop, NAND, ASICINTRODUCTIONA HDD on a host level typical notebook was studied byIlker Tari and Fidan Seza Yalcin [1] using CFD approach.N. Hariharan et al. [2] carried out a similar CFD work effortfor a laptop cooling system with loop heat pipe technology.Qi Wu et al. [3] used HotSpot thermal simulator to carry outfast reasonable accurate thermal simulation based onequivalent circuits of thermal resistances and capacitanceson the SSD drive level. Miaowen Chen et al. [4] have madeefforts on the system level simulation of a tablet withpackage-on-package (PoP) stacking assembly using CFDmethod. Murakami Katsuya et al. [5] showed results oftemperature simulation on NAND flash memory andcontroller which are influenced by their placement. TingYuan Wang and Charlie Chung-Ping Chen [6] studiedtransient thermal simulation of PowerPC with 3-D thermalADI which was a linear-time chip level transient thermalsimulator.Thermal network has a long history of being used fortemperature analysis and is known for its low computationalcost. K. Nishi et al. [6] investigated a tablet device using onedimensional thermal network approach constructed fromthermal conduction simulation results. V. d’Alessandro andN. Rinaldi [7] discussed some issues of the current 2Dthermal models for electro-thermal simulation andinvestigated simulation accuracy.Recently, a network-based approach was developed forthermal simulation of electronic systems [8-10]. Thisapproach decouples the modeling of solid and fluid regions

and integrates them later. The simulation flow consists ofthree steps as illustrated in Fig. 1. First, solid extraction isperformed using an advanced FEM-based numericalalgorithm [8-10] to generate the thermal network of differentsolid components. Note each component from the solidregion such as PC boards or packages can be also extractedseparately. Second, fluid extraction is performed based ontraditional CFD approach to form the fluid network. Third,all sub-networks generated from solid and fluid extractionsalong with the power input are assembled together to formthe complete thermal network of the system. In essence, thenetwork-based thermal model employs a divide-and-conquerapproach for thermal modeling and it features the followingadvantages: First, the model size in network-based approachsignificantly reduces due to the decoupling of solid and fluidmodeling as well as the separate extraction of different solidcomponents. This makes it possible to simulate all detailed3D geometry (traces/vias/etc.) without any simplification.Traditional CFD approach typically ignores such details dueto the tremendous complexity and computational costinvolved in the coupling of entire solid and fluid regions.Instead, it treats complex solid components using effectivethermal properties with a few exceptions such as ANSYSIcepak ECAD option enabling mapped spatially distributedmaterial properties. Second, different solid components canbe extracted separately and therefore each individual solidcomponent is formalized as its inherent thermal network andtherefore forms as standalone module. Consequently, thethermal network of each component is reusable and readilyavailable for various thermal construction and analysis.Third, by taking advantage of the thermal network solver,both steady state and transient thermal simulations becomevery fast.Fig1. Schematic of network thermal simulation flowEXEPERIMENT AND SIMULATION SETUPExperiment setupAs shown in Fig 2, Thermatron S1.2 thermal chamber isused to create a thermally quasi-steady state environment forthe laptop. The laptop platform used in the test is 17” LenovoIdeal Pad 700 model. In Fig 2, the M.2 SSD 2280 modeltested is A400 from SanDisk. Power measurement of keycomponents on the SSD PCB was monitored by connectingan extender with INA231 on 0.01Ohm measurement resistor.Three gauge 32 thermocouples together with Agilent34972A data acquisition system have been used to monitorthe temperature with time: one was attached to the top of oneNAND package (closer to the socket), one was attached onthe platform surface and another was attached on the deviceof copper foil. The SSD will enter the throttling mode usinga combination rule of ASIC and NAND temperatures.Fig 2. Testing platform in a thermal chamberFig 3. SSD test setup inside the chassis of laptopCFD simulationA detailed three-dimensional time-dependent CFDmodel was established in ANSYS Icepak base on theexperimental setup described in the previous session. Twocases have been simulated in this study. A steady state casewas simulated for idle state of SSD and a transient case wassimulated under a typical SSW workload of SSD untilthrottling begins. Both natural and forced convection havebeen included in the simulation. Model layout has beenshown in Fig 4. In the model, mother board was enclosed inthe Chassis with heat convection coefficient assumed to be10W/m2K. Due to the compactness of the notebook, there isnot much air circulation in the chassis and heat is transferredto small side areas of the chassis and the bottom with the helpof thermal dissipation plate and heat pipes [1]. Below thebottom surface heat is only dissipated by conduction throughthe thin air gap which acts more like an insulating layer [1].Double fans have been placed against heat pipe with heatsink assembly. Custom fan curves have been used for bothfans. Grills have been modeled for both inlet and outlet ofthe fan assembly. Boussinesq approximation have been usedunder operating pressure of 101325 Pascal. Ray tracingradiation model has been used. Discretization scheme wasused for the governing equations: standard for pressure, firstorder for momentum and first order for temperature. Heatsources are shown in the Table 1. All power sources are

assumed to be constant except for ASIC power whichlinearly increasing over time as shown in Fig 5. Materialproperties have been listed in Table 2. Chassis outsideconvection heat transfer coefficient was assumed to be 10W/mK with ambient temperature set to 22oC. Radiation hasalso been considered on the chassis.Fig 4 Detailed SSD model in a laptop environmentPMICDDRThermal Power (W)2.50RAMWiFi/BTBatteryHDD3222Table 2. Thermal conductivity in the model (W/mK)MaterialThermal PowerSiliconCopperHeat PipeMother BoardMold CompoundSubstratePolycarbonate ABS148387.640,000(50, 50, 2.5)0.9(46, 46, 2.1)3.11Network Thermal Model simulationIn solid extraction presented here, all solid componentsof laptop were divided into four groups includingmotherboard, SSD, CPU GPU, and RAM. Note that solidcomponents can be grouped in different ways that lead todifferent but equivalent network assembly. Fluid extractionwas performed based on traditional CFD approach aspresented in the CFD Simulation Section. Network modelsof solid components and associated fluid models areconsequently assembled following their original designshown in Fig. 4. Fig. 6 shows the entire thermal modelnetwork assembly for the laptop after solid and fluidextractions. This network assembly is readily used togenerate static and transient thermal solutions with powersources provided for components including CPU, GPU, diesin SSD and RAMs, etc.ControllerNAND2.001.501.000.500.00050Time (s)100Fig 5 SSD major components averaged thermal powerprofilesTable 1. Thermal power sources in the model (W)ComponentThermal PowerCPUGPUSSD45333.7-4.1Fig. 6 Thermal model network assembly for laptopshown in Figs. 3 and 4