More than three years have passed since my last blog about choosing hardware for EDEM. In this time there have been so many rapid changes related to both CPU/GPU hardware and EDEM software that an update is well overdue. Particularly worth mentioning is the introduction of the EDEM GPU Solver, which has made many people ponder how to align their hardware purchasing with future GPU development.
To be honest, there is no unique and ‘right’ answer, but in this post, I will try to provide some recommendations when it comes to choosing a processor (CPU), graphic card (GPU), memory and hard drive. Note that we do not endorse specific providers or vendors, these recommendations are solely based on tests we perform internally in the EDEM Technical Team and are based on our own experience.
CHOOSING THE ‘BRAIN’ OF THE COMPUTER
Choosing a CPU comes down to deciding between two manufacturers: AMD and Intel. Three years ago, I think I gave my clear answer which was Intel. Now with the introduction of the new AMD Ryzen ThreadRipper, the situation is slightly changed. Our engineers report that they can achieve the same high simulation speed by using this new generation of AMD CPU processors.
If you are considering an Intel CPU then the next question is should you pick Intel Core or Intel Xeon? Either will work fine to run EDEM. While Intel Core typically has the higher clock speed and less number of cores, Intel Xeon typically has more processing cores and less clock speed. Previously we recommended Intel Xeon due to the limited number of CPU cores available on Intel Core. However, the situation has changed and Intel Core is now available with more cores and has become our preference for EDEM CPU running. For example, the Intel Core X-series processor family provides 18 cores, and this is more than enough for any user who has up to 18 CPU licences.
According to our performance tests and experiences, EDEM runs well with Hyper-Threading technology, which is a form of simultaneous multithreading technology. When Hyper-Threading is enabled, the number of threads available is doubled. If you have enough CPU licenses, you will see some speedup of your simulation, but the additional cores provided by hyper-threading will not be as good as real cores. When considering your CPU licenses, aim to buy a new CPU with enough real cores rather than using hyper-threading. This will give the best performance and has become possible due to recent CPU development.
In general, there are three factors for choosing CPU, namely number of cores, CPU clock speed and size of cache. The more cores, the more parallelization of the simulation calculation is possible. The higher the clock speed is, the faster the simulation is. As for cache, it is critical to reduce data access time and latency, and a bigger size of cache will result in improved and faster EDEM simulation performance. If you are limited by the hardware budget, you might need to find a compromise among these three requirements.
Graphic cards can typically be classified into three categories: entry-level, mid-range and high-end.
For EDEM displaying, a mid-range graphic card would work fine even for ‘extreme’ cases – say a simulation with 20 million particles. So for most ‘smaller’ simulations, any graphic card will be fine, even entry-level.
20 million particles displayed by AMD Radeon Pro WX 4100
The above figure is a snapshot of EDEM displaying 20 million particles in a mill test case. By using a mid-range graphic card, such as AMD Radeon Pro WX 4100, users can easily and promptly manipulate the viewer, such as rotation, zoom in and zoom out, etc.
Now, it would be a different answer if you wanted to use GPU for your simulations. Initially we did receive some reports from customers saying that they were not achieving the speed up they expected to see for EDEM GPU. However, in most cases the answer was simple, the GPU was simply not good enough. We strongly recommend using a high-end GPU if you want to use the EDEM GPU solver for the simulation.
Both AMD and NVIDIA provide workstation series and desktop series GPU. See the following table:
Shall we choose workstation series or desktop series? The answer is workstation series, namely either AMD Radeon Pro or NVIDIA Quadro series GPUs. In order to achieve high numeric accuracy, the EDEM GPU solver is completely double-precision. Thus, the workstation card which normally have better double-precision capability becomes the natural preference.
According to our performance tests, EDEM GPU has very good scalability, which means that the more processing cores available in the GPU, the faster the simulation speed. Considering EDEM places no usage limit on the number of GPU processing cores, we strongly recommend that users choose high-end GPUs which have the most processing cores. Please note that AMD refers to GPU processing cores as “Stream Processors” while NVIDIA label them as “CUDA Cores”. You can’t really compare the number of stream processors with the number of CUDA cores but you can use these metrics to compare GPUs within each manufacturers range.
In the current market, the top workstation GPUs and their quick specs can be found in the following table:
Please refer to our published benchmark for these three GPU’s performance on EDEM GPU. The GV 100 has the best performance, then the GP100, followed by the WX9100. However, the price of the GV100 and the GP100 is several times that of the WX9100.
In the future, there will be more and more good GPUs released from both NVIDIA and AMD. Please bear in mind, the more processing cores means the faster simulation speed, the more GPU memory means the bigger the simulation which can be run. Always invest in the best high-end GPU you can, our EDEM GPU’s performance is quite linear with the number of processing cores.
Simulation of high shear granulator including 10 million particles
MEMORY AND HARD DRIVE
EDEM is quite memory efficient. For instance, EDEM can run several million particles under 24GB of memory on a desktop machine. Note that sufficient physical memory should be left for running EDEM so don’t load up the machine with other tasks at the same time. If virtual memory is used due to the physical memory being full, EDEM will become very slow. Users should consider expanding the size of physical memory in order to avoid using virtual memory.
Memory frequency is very important. The higher the memory frequency, the faster the memory data access for the EDEM simulation. In general, please choose the biggest memory size and the highest memory frequency based on your hardware budget.
In recent years, SSD drives have become popular. There is no doubt that SSD will greatly improve your simulation by reducing the simulation data file reading/writing time. However, if your simulation has very frequent data saving, you might need a big SSD drive for your computer.
I hope this was helpful. You can check our System Requirements page for more information or get in touch with us for more recommendations.