10 Common Questions about EDEM GPU Answered

David Curry
David Curry
Altair Employee
edited May 28 in Altair HyperWorks

Updated March 2023

In this post we provide answers to the most common questions our EDEM users ask us regarding GPU and multi-GPU. Find out more about the EDEM GPU solver engine, why and when to use GPU and multi-GPU, recommended cards and what speed-up to expect.

  1. What is EDEM GPU and why use it?

The EDEM GPU solver engine allows users to run EDEM simulations using their computers Graphical Processing Unit (GPU). Unlike traditional desktop CPUs, GPUs contain many thousands of compute cores and the Discrete Element Method (DEM) is particularly suited to scaling across these high number of cores. This means the EDEM GPU solver can run simulations faster and also that users can now run larger simulations than what was possible on CPU alone.

  1. What framework is EDEM GPU using?

The EDEM GPU solver is written using the CUDA language, the CUDA solver supports multi-sphere, sphero-cylinder and polyhedral particles.  The CUDA solver runs on NVIDIA hardware only.

There is a deprecated OpenCL solver which runs on hardware from both AMD and NVIDIA however the recommended solver is CUDA.

  1. Is EDEM GPU single or double precision?

There are 3 precision options available:

  • Double Precision
  • Hybrid Precision
  • Single Precision

Check out this blog for further details on precision modes - EDEM CUDA GPU - Precision modes

  1. What cards do you recommend?

We recommend that users choose workstation or data-center grade GPUs for running EDEM simulations.

At the minimum you should have:

  • CUDA enabled GPU with compute capability at least 3.5.
  • Nvidia driver minimum version 11.0 of the CUDA toolkit.

Our recommended GPUs include the following:

  • NVIDIA Volta GPU cards (e.g. V100)
  • NVIDIA Ampere GPU cards (e.g. A100)
  • NVIDIA Hopper GPU cards (e.g. H100)

The range of GPU cards available and the technology behind them is constantly changing and moving fast. Minimum spec cards typically retail for ~USD $1000 and our list of recommended cards go for up to ~USD $11,000 with key differentiators typically being number of cores, size of available memory and FP64 (double precision) performance.

Customers looking to purchase a GPU to run EDEM simulations should check our GPU Benchmark Results which also highlights the specification of cards which have been tested with the latest version of the EDEM GPU solver. In general, cards with a large physical memory, good bandwidth and high FP64 and FP 32 performance perform well when running large EDEM simulations.

If you’re not sure of what hardware you need please contact support and get in touch. We will be happy to talk about results from our most recent rounds of testing. 

  1. Where is the GPU most effective?

In most situations, running simulations on GPU becomes more effective than running on CPU above particle counts of 10k and you start to see significant speed up above 100k. In benchmark simulations containing 1 million particles, simulations run on the Nvidia A100 have shown speed ups of up to 124x when compared against running the same simulation on 32x CPU with the benchmarks showing 187x speed-up on 2 A100 GPU's!

Large simulations that previously would take potentially weeks to run on CPU can now be run in under 24 hours because of the size of the speed up seen between CPU and GPU.

 

image

  1. Do you have examples showing what kind of speed-up can be achieved?

Below are a series of test cases each including 1 million particles. The graph shows the speed up for the A100 card:

Speed up (vs 12x CPU)

image

  1. When to use multi-GPU?

You might consider using EDEM Multi GPU when you need to speed up a large simulation, speed ups of 1.6 - 1.8 times have been seen when going from 1x to 2x GPU on simulations containing 5 to 10 million particles. In addition one of the main benefit of EDEM Multi GPU is when you need to run simulations with a very high number of particles (10’s of millions of particles). This is because beyond a certain threshold a single GPU, with maybe the exception of the newly released 80GB H100 GPU, will not have enough physical memory to run a simulation.

Memory used by the EDEM GPU solver is dependent on many factors including number of particles, number of contact and number of custom properties to name a few. In most cases investing in a more capable single GPU should be considered first before upgrading to multi GPU functionality if the size of your simulation requires it.

  1. Can I run custom physics models (using the EDEM Application Programming Interface) on GPU?

EDEM GPU API, introduced in EDEM 2019, means you can now write and run API models directly on GPU. See our YouTube API overview for more details, also example GPU API models can be accessed in the Altair Community. 

  1. How can I get access to EDEM GPU?

The EDEM GPU solver engine is directly integrated into the EDEM environment and available as standard under Altair Units licensing.

  1. How do I go about using the GPU solver in my simulation?

Setting and running simulations with the EDEM GPU engine is as simple as enabling the GPU engine and then running EDEM the same way you would any other simulation. To get a step-by-step guide check out our eLearning course: Introduction to EDEM which explain how to use the GPU and multi-GPU solver engines.

 

 

 

Tagged:

Categories