FEA simulation, analysis, and virtual prototyping | How we do it 

Hi Ryan... yes, we have a HPC pack allowing for use of a GPU with a limited number of cores. Although, I believe we have trialled the use of GPU's on some jobs and often it hasnt resulted in significant performance improvements for a lot of our static or quasi-static FEA analyses. We have successfully used gaming PC's due to their high single core throughput when using a limited number of cores. We are now starting to ramp up usage of the Ansys 'cloud licenses' (I forget the name of the new system) which is great when we have more work than our perpetual licenses will permit. We have not completed any back to back comparison of Xeon performance with the new instruction set. This would be good to see any data on the performance improvement per clock cycle and whether this would exceed the throughput from a higher clocked (overclocked) gaming PC. Classicaly, we have run the gaming PC's because of ease of access to parts if something fails, the much lower initial cost of the machine, and the high clock speed/single core throughput using our perpetual licenses. But this may need to be reassessed as we scale more. Im sure Ansys, or your VAR could provide acess to info and even machines to test the performance differences. I'd be super interested to undertake that process. Maybe for another video! :) Let me know your thoughts.

AMD's 7000 series also now support the AVX512 instruction set. The main issue with gaming chips over hedt/HPC chips is lack of memory bandwidth, especially when using more than 6-8 cores in dual channel as you become I/o bottlenecked pretty quick. The main discrepancy now between intel and AMD is the lack of support of the intel maths kernel libraries with AMD chips as it's well regarded as the best optimised for the required matrix calcs. I'd be interested to see a benchmark from Ansys on Intel's mkl Vs AMD blis to see how much of a difference there is.