Cliquez pour modifier le style du titre
Transcription
Cliquez pour modifier le style du titre
Thermal Modeling Methodology for Fast and Accurate System-Level Analysis: Application to a Memory-on-Logic 3D Circuit Cristiano Santos1,2, Pascal Vivet1, Philippe Garrault3, Nicolas Peltier3, Sylvian Kaiser3 1CEA-LETI, FR 2UFRGS, BR 3DOCEA Power, FR www.cea.fr & Thermal issues in modern SoCs Cliquez pour modifier le style du titre Increasing thermal issues Technology scaling = ↑higher power density 3D stacking with TSVs = ↑↑even higher power density and ↓reduced heat dissipation properties Higher lateral thermal resistance due to thinned wafers Poor conductive materials used to bond stacked dies Temperature impacts Power consumption Peak performance Ageing Package costs How to perform thermal modeling to enable early exploration of thermal issues including 3D? & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |2 Thermal Analysis: state of the art Cliquez pour modifier le style du titre Low-level tools: Multiphysics simulation : FloTherm (Mentor Graphics), Icepack (ANSYS), Marc (MSC) , COMSOL General-purpose FEM solutions with no specific support for IC design flows Post-layout level : HeatWave (Gradient DA), Apache (ANSYS) Adapted for short term analysis at die level but long simulation times for multiple power scenarios Those solutions are time consuming methods thus not suitable for system-level simulations High-level tools: Architectural and system-level thermal simulators must support: Fast transient thermal analysis to be used in power-thermal coupled simulations Different granularity scales: from large structures (package, interposer and board) to very fine-grain elements (TSVs, C4-bumps and copper pillars) with both high impact on model accuracy Existing solutions: Mostly academic tools like HotSpot (Virginia) and 3D-ICE (EPFL) No support for fine-grain structures with heterogeneous distribution No solution for fast transient thermal simulation of complex systems with unrestricted support for fine grain structures required for 3D integration, flip-chip designs or BGA-like packaged circuits & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |3 Outline Cliquez pour modifier le style du titre Introduction State of the Art Thermal Modeling Methodology using ATM WIOMING : a Memory-on-Logic 3D Circuit Correlation of Thermal model wrt Silicon Measurements System Level Exploration of Thermal Impact of 3D Conclusion & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |4 Thermal Modeling Methodology Cliquez pour modifier le style du titre Thermal modeling based on DOCEA™ ATM tool: Based on a numerical finite difference method Heat transfer modeled via full 3D heat diffusion in solid materials with no restriction to heat flow paths Highly compacted thermal model (CTM) with good tradeoff between accuracy and efficiency API + GUI for easy model creation/updates CTM generation fully automated by Python scripts Four main contributions: Automation of thermal model construction by parsing Design Floorplan (LEF/DEF format) Automated homogenization methodology to reduce complexity while preserving accuracy Validation of proposed approach on a Memory-on-Logic 3D circuit System Level Exploration of 3D Thermal effects & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |5 Thermal Model : input data Cliquez pour modifier le style du titre LEF/DEF parser implemented within ATM Die size and placement of power sources, TSVs, u-bumps and C4-bumps can be read into ATM from initial explorative floorplans or final layout databases Input data organized as follows: Material thermal property library: CSV file Circuit description: parsed from floorplan (LEF/DEF format) and saved as CSV files Technology settings: thickness, diameter and material used for every structure of a specific technology CSV format is widely used for system-level exploration Allow fast iteration and manual modifications & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |6 Thermal Model: Material Homogenization Cliquez pour modifier le style du titre Material Homogenization: This technique consists in calculating the equivalent anisotropic material properties for a heterogeneous layer containing multiple structures A set of geometries are identified to go through material homogenization Once the equivalent material is calculated, those original geometries are then replaced by simplified structures Package solder balls Array of TSVs and u-bumps Material homogenization applied to : Layer stacking structures + Array areas Objective is to reduce the number of geometries going for extraction while keeping the spatial accuracy Fully Automated using Python API & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 |7 WIOMING, a Memory-on-Logic 3D Circuit Cliquez pour modifier le style du titre WIOMING circuit floorplan D. Dutoit, et al. "A 0.9 pJ/bit, 12.8 GByte/s WideIO Memory Interface in a 3D-IC NoC-based MPSoC“, VLSI-Symposium, 2013. WIOMING circuit Many-core architecture, STMicroelectronics 65nm TSV middle (diam 10µm), µ-bumps (diam 20um) 3D Assembly : Die2Wafer, Face2Back, FlipChip Stacking a WideIO compatible DRAM & © CEA. All rights reserved WIOMING circuit instrumented with 8 Heaters [Can generate each 1Watt] 7 Thermal Sensors [Accuracy ~1°C after calibration] Full Thermal Software Control on chip & board to perform accurate 3D Thermal Characterization P. Vivet | DAC’14 User Track | 2-5 June 14 |8 WIOMING ATM model: from board to chip level Cliquez pour modifier le style du titre +6K circuit structures Board-level including PCB, socket and package Chip-level with stacked dies, TSVs and u-bumps & © CEA. All rights reserved 27 power sources 26 areas for heat exchange Multi-corner CTM generation +2k lines of Python script P. Vivet | DAC’14 User Track | 2-5 June 14 |9 Model Compaction Results & Tool Performance Results Cliquez pour modifier le style du titre Material Homogenization & Model Reduction Results 130x less geometries after material homogenization 60x less nodes after material homogenization 570x less nodes after model reduction # geometries # defined materials # extracted nodes System physical representation Before After After homogenization homogenization reduction +6k 71 -23 44 -18 million 231k 405 ATM Tool Performances a Per simulation time step Highly compacted thermal model for static and dynamic thermal analysis Very fast transient simulations compatible with system-level exploration & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 | 10 WIOMING : Thermal Correlation Cliquez pour modifier le style du titre Simulation vs. Silicon comparison for various profiles : Steady-state analysis: Very good accuracy of hot spot evaluation (avg=3.96% and worst=13.41%) CTM is able to capture the spatial temperature distribution Transient response (staircase stimuli) Temp. vs Power for various hot spot distances High thermal time constant due to package socket and board Good fitting between simulation and measurement data plots Thermal time constant error in the 5% – 40% range Transient response (PWM shape) Good fitting between simulation and measurement data CTM able to model the multiple thermal time constants of the system Transient step response PWM response for various PWM periods & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 | 11 Accuracy Impact of Material Homogenization Cliquez pour modifier le style du titre Steady-State analysis experiments with: 1) No homogenization procedures (all fine-grain structures are ignored) 2) Homogenization without selecting localized areas (flat) Full homogenization approach (reference) Case 1 -> average error = 51% | worst case error = 81% Case 2 -> average error = 7% | worst case error = 14% Properly selecting regions to apply material homogenization brings considerable accuracy to the spatial temperature distribution No Homogenization & © CEA. All rights reserved Flat Homogenization P. Vivet | DAC’14 User Track | 2-5 June 14 | 12 Impactpour of board components Cliquez modifier le styleon dusilicon titre Not accounting for the components mounted on the board introduces an additional average error of 24% in the steady-state temperature. Worst case error is 43.66%. Those results show that : Poor power modeling of the board elements will have strong accuracy impact ! Thermal modeling approach being focused either only on board-package or silicon level is not enough for accurate thermal analysis. Error when removing PCB components & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 | 13 Exploration Thermal Impactle ofstyle Die Thickness Cliquezofpour modifier du titre Compared 3D die version with 2D die versions For various 2D die thickness (200/300/500µm) For same hotspot power budget (4 Watt HotSpot in die bottom left) ? 3D die temperature increase versus 2D die versions (wrt hotspot distance) Very strong effect of die thickness : o Thin 2D die has worse behavior than 3D die o Thick 2D die provides better dissipation & © CEA. All rights reserved Industry is driven by reduced form factors Beware of thermal effects ! P. Vivet | DAC’14 User Track | 2-5 June 14 | 14 Exploration ofmodifier Thermal le Impact of TSVs Cliquez pour style du titre TSV Thermal Properties : TSV Ø10µm, pitch 40µm, SiO2 isolation layer 0.25µm Thermal Conductivity of a TSV array (Homogenized material) Thermal Conductivity Thermal Impact of TSV Kxy = 133 increased horizontal resistance negative impact for hotspots Kz = 159 reduced vertical resistance good for average power * Thermal Conductivity of silicon only : 150 W/mK Exploration of WIOMING Circuit Testcase : Comparing with TSVs vs. without TSVs Power dissipation Hotspot – a) Hotspot – b) Uniform Temperature difference + 2.22% + 5.89% - 0.02% TSV does not reduce temperature for uniform power Due to very low TSV density (3%) But may lead to temperature increase for hotspot & Must take care of TSV Hotspot placement !!! © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 | 15 Conclusion Cliquez pour modifier le style du titre Thermal Model in ATM tool can be fully automated DOCEA™ ATM tool providing an API for Python scripting Die Floorplan parsed from LEF/DEF database Material homogenization, to reduce model complexity, maintain accuracy Thermal correlation with a Memory-on-Logic 65nm 3D circuit Steady State analysis : less than 4% error, accurate evaluation of Hotspots Transient step response : 5% – 40% error in thermal response time System-level thermal model: accuracy versus simulation time Must focus on all levels: from board & package to die fine grain structures Efficient compaction engine allowing fast thermal simulation and system-level exploration System Level Exploration Examples : Die thickness impact on temperature : 2D very thin die can be worse than 3D die TSV impact on temperature : care must be taken for Hotspot & TSV placement Next steps Parameter sensitivity analysis for enhanced early thermal exploration Support for 3D stacking configuration files and std package formats & © CEA. All rights reserved P. Vivet | DAC’14 User Track | 2-5 June 14 | 16