Author Topic: Another Threadripper thread...  (Read 2567 times)

0 Members and 1 Guest are viewing this topic.

Offline alexalexalex

Another Threadripper thread...
« on: November 29, 2018, 08:31:56 am »
Its been a few months now, is anyone running the threadripper 2990wx ?  I am buying new machines for my ID department and would like to know people's experiences.  I feel the 2990wx is good because of the cost and core count.  We use Keyshot and Solidworks. 

Offline mattjgerard

Re: Another Threadripper thread...
« Reply #1 on: November 29, 2018, 10:09:29 am »
Posted a couple right below this one-

Offline alexalexalex

Re: Another Threadripper thread...
« Reply #2 on: November 30, 2018, 08:20:55 am »
Thanks, Yea Iv'e read through all the relevant threads on here, and was just wondering if more people had used it or could explain this article where the bathroom render was drastically poorer performance.

Offline KeyShot

Re: Another Threadripper thread...
« Reply #3 on: November 30, 2018, 07:23:48 pm »
We are in the process of investigating the performance on the threadripper (32/64T). The performance is great with product mode, but it may be that interior mode is affected by the fact that the threadripper is a numa architecture. We will keep you posted.

Offline Tharic-Nar

Re: Another Threadripper thread...
« Reply #4 on: February 01, 2019, 09:31:33 am »
I work with Rob at Techgage, the author of the article that highlighted the issue with the Interior render. KeyShot may have figured out the issue already with the 2990WX, but I am here to make people aware of some of the extra testing we did on the CPU:

The short of it is that Windows' thread scheduler is not optimized currently, causing issues with cores without direct memory access. AMD Dynamic Local Mode tool does not solve this issue for KeyShot, however a third-party tool called CorePrio does help in some cases. When enabled, it nearly doubles performance - after a few attempts. When doing the render passes, CorePrio did not work on the first attempt, but on subsequent attempts, it did resolve the performance regression.

When we tested applications under Linux, we saw little in the way of regression, so it's definitely a Windows related issue for the most part.