Ansys2025全球仿真大会：Ansys CFD 2025新功能介绍

AnsysCFD

2025新功能介绍刘文东©2025ANSYS,

Inc.原生GPU物理模型&性能增强•FGM

Combustion•FW-HAcoustics•

SurfacetoSurface

(S2S)

Radiation•Discrete

Phase

Model(DPM)•Multi-phaseVolumeof

Fluid

(VOF)•Compressible

liquid•

耦合求解器计算时间缩减30%,内存消耗降低20-25%•多显卡的扩展性能增强•

GPU后处理收益o能源&动力行业涉及的燃烧问题o旋转机械噪声o

电子散热o

多相流问题o

水锤...

工程应用解决方案更新

•

Tsinghua’s

N-Equation电池包热失控模型•

均温板仿真•Freeflow电机&齿轮箱甩油&水管理

Benefits前处理及用户体验•Case间设置复制&粘贴&对比•

Pyfluent显示增强•

嵌入oSL后处理FMU输出功能•

Lee&SMB模型增强•

Fluent

Meshing新功能•

RapidOctree•

usd格式支持•

加速ui显示收益o

性能优化o

易用性提升o

跨平台协同o

自动化和仿真精度•

电池包热失控•整车外气动&热管理前处理•水管理&电机、齿轮箱散热润滑•

VC散热仿真powering

InnovationThatDrivesHnnanAdvancennentAnsysCFD

2025

HighlightsFluent原生GPU求解器功能更新©2025ANSYS,Inc

.4

更快的设计迭代•相比CPU拥有更多的ALUs•更低的能耗•更高带宽的内存•更高的双精度浮点数•

相当的时间内获取更高的仿真真实性•

or:

获取相同的仿真洞察力需要更少的时间•

or:

相同的时间完成更多的设计验证GHz,

40

coreSKX

2.4GHz,

40

coreBDW

2.3Ghz,

36

coreHSW

2.6GHz,

28

coreIVB

2.7GHz,

IVB

2.7GHz,

24

core24

coreSingle

H100~9Xgreater

than

112

CPU

cores20082009201020112012201320142015201620172018201920202021202220232023Sedan_4m13446101013142020262838417091GPUs助力下一代CFD91X2.8GHz,WSM2.93GHzW,SM

2.93

GHz,

16

corecore

12core

12

coreNHM

2.8

GHz,WSM

2.93GHzW,SM

2.93GHSz,NB

2.60

GHz,8

core

12core

12

core

16

core20092010

2011

20123

4

4

6©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennentHSW2.6GHz,BDW

2.3Ghz,

28

core

36

core2015

201613

14Sedan_4m单个双插槽节点上的归一化求解器效率性能数据由英特尔提供20192020

2021

2022

2023

2023262838417091IVB

2.7GHz,

IVB

2.7GHz,

24

core24

core2013

201410

10SPR2.0

GHz,112

core

ICX

2.3GHz,80

coreCLX2.4

GHz,

CLX

2.9

GHz,

48core

48

core~630XHTN3.0GHz,

8

coreHTN3.0GHz,

8

core20082017

201820

20SPR

MaxCPU

1.9

GHz,112

coreSPR

Max

CPU

1.9GHz,

…SPR2.0

GHz,

112

coreSKX

2.4GHz,

40

coreSKX

2.4GHz,

40

coreCLX

2.4

GHz,

48

coreCLX

2.9

GHz,

48

coreICX

2.6GHz,

64

coreICX

2.3GHz,

80

coreICX

2.6GHz,

64

coreSedan_4m

1SNB

2.60GHz,NHM8GPU显卡硬件信息：

NVIDIAA100-SXM4-80GB104SM测试案例信息：uDrivAer250M+LES加速，以

112个CPU核心（8336C）为参照；u时间步长5e-5,内迭代次数4次，仿真时间1s

，8卡耗时5h；

结果总结：l

一个80GB的A100最大可替代约

1260个CPU核心；l

3.4GHzCPU计算速度为2.3GHzCPU计算速度的1.5倍，计算速度

比约等于主频比；l

双精度计算速度为单精度的65%；l

单精度需要的GPU内存，每1M网格需要1G内存；l

双精度需要的GPU内存，每1M网格需要1.5G；P。werinnn。vatienThatDrves

HnnanAdvancennentnsysNormallize250.0CPU型号主频/睿频单节点核数Intel(R)Xeon(R)

Platinum8336C2.3/3GHz112Intel(R)Xeon(R)

Gold6467C3.4/5GHz120200.0149.9150.0121.9DrivAer外气动测试结果251.8164.2127.645.01.0

1.6

1.2

1.8

3.55©2025ANSYS,

Inc./

Proprietary.

Do

NotShare.100.050.00.094.976.281.450.6205.30.0Equivalent

CPU

Ave

IterationTime16000.814000.712000.610000.58000.46000.34000.22000.1002GPUs4GPUs6GPUs8GPUsTimeCPUS

线性(CPUS)6/nsys

powering1.

Without

solidand

energymodel;2.

half

modelwithoilthrowingandspraying;3.

8GPUs

on

NVIDIA

RTX5880Ada;4.

VOF

model+

pseudo-

transientsolver;5.

233

revs(2s)tosteady

result.6.

25M

cells&1e-3sand

16.6Mcells&1e-4stimesteptest;Water-cooledmotorwastested,used22.6

millionpolyhedronmeshand

iterated

1,000timesinjust

13minutestoobtainthefinal

result.Air-cooledmotortest,with44

million

polyhedronmesh,ittookonly

15

minutestoobtainthefinal

resultafter

1,000

iterations.25.0MCells;Pseudotimestep=1e-3;Physicaltime=2.0s;Totalwall-clocktime:22minutesEmotorVOFGPUteston

NVIDIA

RTX

5880

Ada16.6MCells;Pseudotimestep=1e-4;Physicaltime=2.0s;novationThatDrivesHnnanAdvancennentTotalwall-clocktime:140minutesInEquivalent

CPU.0.220.2277.5X65.3X46.1X35.3X34.0X13.5X17.6X13.3X.17.6X

9

4X12.6X18.7X9X..MIUPS86311248481598225429394164494360211248021541629.85

1191.2

2170.285%73%100%94%89%87%82%73%100%93%100%96%100%95%86%25

1XElectricity

Cost

of

Job*$5

$3.89

$4

$3

$2

$1

$0.23

12288xCPU8xA100•Job

cost

=

w

TDP

tE-w=computeworkload

(assume

90%

capacity)-TDP=thermal

design

power

(kW)-t=

run

time

(hr)-

E=costof

electricity

(assume

$0.13/kWh)$350,000

$300,000

$250,000

$200,000

$150,000

$100,000

$50,000

$0

12288xCPU8xA100-time

HW

Cost(processors/cards

only)7

©2025

ANSYS,

Inc.

*Doesnotaccount

foradditionalcost

topowerthe

coolingsystemCPUsVS.

GPUs$0

Monthly

Electricity

Bill*powering

InnovationThatDrivesHnnanAdvancennent$5,000$4,000$3,000$2,000$1,000$-4A100

8A100

GPUs

GPUs8L40GPUs8H100GPUs2H200GPUs8H200GPUs4H100GPUs4H200GPUs$2708xA10048

L40

GPUs32

L40

GPUs64

L40

GPUs24

L40

GPUs16

L40

GPUsOne12288x

CPU1228824.2X$4,52940008192500060000.160.400.210.120.290.290.160.110.090.060.050.410.319由包含CUDA

核心的流

由包含流处理器的计算多处理器(SM)

组成

单元(CU)

组成•

GPU求解器需要的HPC授权数量由显卡计算单元(graphics

compute

units=GCUs)

决定-GCU=SM

(if

using

NVIDIA)or

CU

(if

using

AMD)•

GPU内部包含多个

GCUs,每个显卡计算单元又包含多个逻辑运算单元(ALU)•

越强大的GPU卡通常拥有越多的

GCUs-L40

vs

A100

例外5.04.43.7

3.8

3.22.3

2.3

1.91.21.0到nvIDIA.Licensing–

基本硬件信息Fluent

Car

2M

Octree,

Flow+Turbulence,SpeedupsingleGPUvs

1

CPU

node©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennentExampleFluentoutput

whenrunningAMD:8xMI210,104

CUs

eachAMDMilan128coresGCUsRTX

RTX

AMD6000

A6000

MI210RTX6000Ada142L40142A10040GBA10080GBH10080GB8

P100104108108132845672Licensing–CFD

Enterprise&Ansys

HPC•

CFD

Enterprise含有：40GCUs•额外所需要的GCUs授权通过Ansys

HPC/Packs/Workgroups获取-1

HPC

task

=

1

additional

GCU-HPC

Packs的计算方法不变(e.g.

3

Packs

enables

128

additional

GCUs)Note:•GPU求解器的前后处理过程并不占用HPC授权，占用的HPC数量与打开时使用的CPU数量相同•授权是基于所有显卡的GCU数目，与显卡的个数无关•

A100GPUcardcontains

108

GCUs•

4xA100GPUcards

requires:

(4

*

108)-40

=

392

HPCtasks•

4

HPC

Packsor392Workgrouptasks#

SMsHPCWorkgroupHPC

Packs1–

400041–

481–

8149–

729–

32273–

16833–

1283169–

552129–

5124553–

2088513–

20485•启动面板和信息输出面板会显示可占用的GCUs使用信息•一张GPU卡的GCUs会一次性全部使用，而不能仅使用其中的一部分©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennentDisplaywill

be

updatedfor

25R2EXAMPLE9

©2025ANSYS,

Incion

a

riveserinInnova10u

an

vancennen

单一的授权允许使用无限数量的CPUs

或者GPUsCPUs

GPUs

Parametric•2月份发布•不需要额外的HPC授权-

E.g.

runasingle

Fluentjobon

1GPUor

10

GPUs

with

the

same

license•兼容optiSLang或不同HPC数量运行Ansys

CFD

HPC

UltimateCFD

without

HPC

limitsReach

outto

learnmore!

/FW-H声学模型•

“on-the-fly”模式,导出麦克风压力时域信息•“Read-and-compute”与CPU方法流程一致250200150100500214.9164.3115.712.08

23.08128

CPUs

256

CPUs

4

nodes8

nodesRe=0.9.105,

M=0.2FW-H

integrationsurface:

internalzone

and

wall1

Nvidia®A100GPU~

640

CPUcores

Intel®Xeon®Gold6242Targetapplications:o

Turbomachinery

bladeso

Jet

noiseo

FanacousticoNoise

infree-flowconditionsFlow

pass

cylinder

acoustics8m,

hex,sp,

SBES,FW-H©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent-

导出

asd或者ard文件dp/dt,far-field

BC2GPUs4GPUs

8

GPUs

CPU

A10011

MIUPSS2S辐射模型与滑移网格-

基于GPU的S2S辐射模型-

每次网格更新时，都会自动重新计算角系数。-

GPU求解器中的计算速度比

CPU求解器快得多，这使得这种方法在许多情况下都可行。©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent12

•密度为压力的函数(Tait’s

law)•用于压缩和膨胀•可用于实际物理问题或者提高数值稳定性•当前版本限制:-

仅支持单相流Water

HammerPipeflowwith

steadystate

profiles

issuddenly

closed

at

one

end,which

produceswater

hammereffect.CPUvsGPUcomparisonTargetapplications:o

水锤及水下噪声o

数值稳定性调整©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent可压缩液体Areaweightedaverage

pressureatthe

outlet13

powering

InnovationThatDrivesHnnanAdvancennent•组分输运方程的源项•

由颗粒到混合气体的传质支持蒸发和沸腾DPM:传质模型14©2025ANSYS,

Inc.AirWaterPaste15

©2025

P。wering

Inn。vationVOF模型•兼容:-

密度变化-

Steady&transient-

MRF-

Sliding

Mesh-

多组分模型-

Non-Newtonianflow•

Carreau

and

power

law•

支持表面张力系数指定•

壁面附着(润湿模型)，指定壁面接触角ThatDrivesHnnanAdvancennent42.0422.2814.019.685.61注液,VOF+表面张力模型+1M

poly,doubleprecision,

MIUPSCPUIntel(R)

Xeon(R)

Gold

6242GPU

A10080GB2xA100GPUs~10xfasterthan

128CPU

coresVOF冷却液加注©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennentExcellent

performanceandscalability73.6858.18128cores256cores512cores1GPU2

GPUs4GPUs8GPUs16

•

浸入过程涉及空气被电解槽液排出。在金属表面内形成的气穴会导致接触不良和涂层工艺不良。•

浸出过程涉及从车身排出电解液。任何残留液体进入下一阶段的喷漆工艺都可能导致局部污染和质量问题。•GPU+VOF+

Dynamic

Mesh。资源消耗：200s,NVIDIA

L40*4块*22小时©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent电镀排气17

•18©2025ANSYS,

Inc.

Expressions表达式(β)•2025

R1

引入：•

报告中的表达式作为

monitoring,

plots•

参考坐标系随时间变化的转速使用Expressionsoerin

InnovationThatDrivesHuanvanceent•On

Nvidia

hardware,CUDA

12.8

must

be

explicitly

installed

at

thesystem

levelto

use

Expressions

in

Fluent-Installingthe

GPU

driver

by

itself

is

not

enough-Refertothe

Nvidiawebsitefor

download

and

installation

instructions-Only

GPUsthat

arecompatiblewith

CUDA

12.8

can

be

usedNoadditional

installation

neededforAMDGPUs

on

Linux•Maxwell

(2014)

microarchitectureand

newer•支持“结果光顺”,根据边界或单元的中心插值-

通过GPU原生实现该过程•平滑显示可以更好地判断梯度•提高曲面上的显示质量•支持XY-PlotGPU

Direct

Post

Processing

[β]Facevalues

(top)

and

nodalvalues(below)

on

planesurfaces

in

Direct

Post©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent19

前处理及用户体验更新©2025ANSYS,

Inc.•WTM在FluentWeb

UI模式可用-

诊断选项在右下角快速仿真工具-

通过右键即可在目录树直接交互式设置©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennentWTM–Web

UI21•在Watertight

网格划分工作流程中灵活地进行分层薄网格划分-自动检测并排序合适的局部区域和源/目标-

在非薄区域自动过渡到非结构化-

导入网格格式时，需要在“ManageZones

”

对区域进行区分WTM–

自动化薄体网格©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent22WTM–周期性对间拉伸•一次操作/任务即可拉伸多个不同尺寸的管路网格•完全共形网格powering

InnovationThatDrivesHnnanAdvancennent23

©2025ANSYS,

Inc.Rapid

Octree

(top-down)Rapid-Octree快速八叉树网格-

“自顶向下”的网格技术，基于CAD直接生成体网格，显著减少网格生成时间-

可容忍低分辨率CAD，无需过多预处理-

快速、近乎线性的并行效率-

高度自动化，几乎不需人工干预1024256

cores,

336s●

2048

cores,

103scores,

147sMesh

refinement

regionsWorkdirectly

ontriangulated

CAD•

Cores

used

:

288•

Volume

meshingtime

:

12mins•

Cell

count

:

182millions•

Quality

:0.018 PerfectScalingOctree-GenerateMesh24

©2025ANSYS,

Inc.Zonal+Angular

RefinementsNon-conformal

multivolume100001000100Prismatic

LayersDrivAer

1.7BCells-Time

Scalingpowering

InnovationThatDrivesHnnanAdvancennent100001000100oering

Innova

ion

a

rives

u

an

vance

en25RO+GPUAppliedto

LESof

an

Aircraft

High

LiftWorkshop

5:GPUsolverforAircraftsimulation

using

LESand

Rapid

Octree

mesh1

BillionCellsGeneratedin28

minon

512

cores26

©2025ANSYS,FTM

LeakshieldTask

(hidden)•25R2

FTM流程可使用Leakshield-

基于泄漏路径封闭几何-

配合RO网格划分方法powering

InnovationThatDrivesHnnanAdvancennent•

Fluent

中嵌入式

oSL参数化研究后处理-

参数图-

蚁丘图-

历史记录-平行图•

直接从

FluentAMOP(Adaptive

Metamodel

ofOptimal

Prognosis)导出FMU格式-可用于系统级仿真工具（例如

Twin

Builder）-避免在

optiSLang

中创建

FMU

的额外后处理在fluent中optiSLang后处理功能©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent27

©2025ANSYS,

Inc.

。

ering

Inn。vaionarivesuanvanceentUniversalScene

Description

(USD)

Export•输出通用场景描述格式USD•USD

旨在实现可视化，可在

OmniverseViewer或

Blender

等许多第三方应用程序中打

开。28工程应用解决方案更新©2025ANSYS,

Inc.Tsinghua’s

N-Equation电池热失控模型•热失控模型框架-

相对于现有的一方程和四方程普适性和灵活性更强-

可通过DSC实验拟合多个Arrhenius反应方程，精度更高©2025

ANSYS,

Inc.

powering

InnovationThatDrivesHnnanAdvancennent30热的过程•相变发生在交界面wick

and

vapor-Flat

Interface

Kinetic

Theory•相变发生在

wick

and

vapor交界面一定厚度内-

Lee

model-Needs

Ghost

Fluid

Method•单相流方法-Vapor和wick分为两个单独的计算域•两个计算域分别模拟蒸汽和水,通过添加源项的方式模拟相变和传Tsat•

Zonal2-phase

approach-

Flow

in

porouswick

isa

2-phase

flowwhileflow

invapor

core

is

asingle-phaseflowLee

modelbased

numericalschemeforsteadyvaporchambersimulations,Yusuf

Rahmatullah,Tsrong-YiWen,

InternationalJournalof

Heatand

MassTransfer201

(2023)

123636T均温板(VC)CFD仿真现状Only

2-phaseflow

inwick

region.Single

phaseflow

invaporcoreNo

benefittotiltangle

analysisIn-house

CodeA

heat

pipecalculation

method

based

onthetwo-phase

mixture

modelof

porous

medium,

BoXu,et

al,

International

Journal

of

Heat

and

Mass

Transfer

238(2025)

126447•

Full2-phaseapproach-2-phase

flow

both

in

wickandvapor

coreCondensation/Evaporatio

n

happens

hereVapor

pressure

here

is

usedto

calculatethesaturationtemperature

ofva