Dear all,
I have started this thread and I am hoping to help you with my answer. It is based on my experience with our new simulation computer (configuration can be found at the beginning of the thread) and the contact with Autodesk and MF Software (the redistributor of Autodesk Moldflow in Germany). MF Software was very helpful on this matter.
First you have to think about what kind of analyses you will run and how you use Moldflow in your daily work:
- The filling, packing, warping analyses require much computing power, but not so much RAM. This means for that kind of analyses you need a powerful CPU, but not so much RAM. I was not even able to fill 16 GB of RAM on our old workstation with 2-3 filling jobs at the same time.
- The cooling simulation requires much RAM, but not so much computing power. If you run several cooling simulations at the same time, I recommend 64 GB of RAM. I managed to almost fill this amount entirely with 3-4 cooling jobs at the same time.
According to MF Software and also confirmed by my experience, it does not makes sense to use CPUs with more than 6-8 physical cores. I think Autodesk does not mention any limitations for Moldflow, but I think one can say for certain that Moldflow supports multi-core computation. However, it seems that even when I run 8-10 jobs at the same time on our new workstation, Moldflow does not really take full advantage of all the cores (physical and virtual). When looking at the Windows ressource monitor, some of the cores are hardly used and run only at very low levels of usage.
If you are thinking about buying a new computer, I recommend therefore that you go for a CPU with about 6-8 cores and as many GHz as possible. That's why I have chosen the XEON E5-2687W v2 with 3.4 GHz (Turbo: 4 GHz). It provides simply the highest frequency I could find for the right amount of cores. In addition to the frequency, my feeling is that the cache per core is also very important to keep the CPU at high pace.
"So you're telling that a Dual Xeon E52620v3 could have worse performances than a single E3-1240v2 or a i7-4790k for example?"
According to the specs of Intel, the Xeon E5-2620v3 has 6 cores and 2 GHz basic/2.5 GHz turbo frequency. The E3-1240v2 has 4 cores and runs at 3.4 GHz basic /3.8 GHz turbo frequency. The i7 has also 4 cores, but runs at 4 GHz basic/4.4 GHz turbo frequency. If you run only 1-2 jobs on your machine, I believe the i7 would be the fastest and the E5 the slowest CPU.
In my opinion one should consequently also take into considerations how many jobs run on the workstation at the same time. I believe, if you work in more "sequential" manner, meaning you run one analysis, look at the results, change your model, run the analysis again and iterate that process many times, the better choice is a very fast CPU with smaller amount of cores, like the aforementioned i7.
Vice versa, if you work in a "parallel" manner, meaning you built up the simulation model and then want to run the same model let's say with different materials, you want to analyze them at the same time and compare the results rather than analyze one after the other. I think it makes then more sense to get a new machine with 6-8 cores and perhaps decrease the speed for getting the most value for your money. In your example this would thus be the E5.
In my experience, most ATI/AMD graphics cards are not supported by Autodesk Moldflow. They cannot be used for computation support of the CPU. On our old workstation the ATI graphics card did cause a lot of trouble. I therefore strongly recommend NVIDIA cards.
According to my talk with MF Software, you can say the GPU computation speed is 8x of a CPU as a rule of thumb. I guess that makes sense if you look at the pure number for Gigaflops (floating point operations) per second and I am pretty sure you can find confirming opinions somewhere on the internet. I can personally also confirm that the GPU is much faster than the CPU. The NVIDIA K500 in our new workstation can be used by Moldflow. When looking in the analysis log file while the analysis is running, the time steps are worked off and coughed out very quickly. As soon as the message appears the GPU does not have enough RAM anymore and is switched off for the computation, the working off of the time steps significantly slowes down. However, this is a personal perception and not based on any serious analysis. As just mentioned, the GPU is useful as long as the integrated RAM of the graphics card is large enough to accomodate the simulation model. With every time step and therefore every node that the plastic flow reaches my impression is that the model becomes larger. For large models and a small amount of RAM, the model is eventually too large for the graphics card.
That is the reason why we have chosen the K5000 with 4 GB RAM instead of a combination of the K4000 and the Tesla K20. The Tesla K20 might be much faster compared to the K5000, when you look at the floating point operations. Since the RAM is much smaller, it will be switched off much earlier in any analysis though.
To sum it up for a fixed budget:
- Few jobs: CPU as fast as possible, as much cache per core as possible, 4-6 cores.
- Many jobs in parallel: 6-8 cores, as much cache per core as possible, CPU bit slower to stay in budget
- Many cooling jobs and/or no mesh aggregation: much RAM (>= 64 GB)
- Focus on filling/packing/warpage and/or mesh aggregation on: less RAM (<= 32 GB)
- NVIDIA GPU with much RAM to take advantage of it as long as possible.
For your information, the default setting for any analysis is mesh aggregation being switched on.
Hoping this helps!
Max
![Autodesk Design & Make Report](https://forums.autodesk.com/t5/image/serverpage/image-id/1350519iCB55845EBFC43016/image-size/large?v=v2&px=999" "Autodesk Design & Make Report")