[Kevin.B]

The company I work for is weighing their options for when they buy Ansys to do turbomachinery design. We are hoping to get as many turbine calculations in a day as possible. We have two computers that we think might work, but do not know which one to use, or if we need to build a much more powerful workstation.

Option 1:
i9-9900k
4x16gb dimms of DDR4-2666
Quadro P1000
Windows 10


Option 2:
Dual socket E5-2630v3
8x8gb dimms of DDR4-2133 ECC
Quadro P1000
Windows 10


If neither one of these will be fast enough, would an AMD Threadripper or Epyc based system be better? Or a newer Xeon base system.

[flotus1]

If neither one of these will be fast enough

Nobody knows what "fast enough" means for you. For me personally, there is no fast enough when it comes to computers.
All I can say with confidence is that the Xeon system will be faster. If building an entirely different system is worth it depends on the price you would pay for the xeon system, and how much you would be willing to pay for an even faster machine. AMD Threadripper is not an option here because it is slower.

[Simbelmynë]

I agree. If you wait until the new rumored generation of Threadrippers release then it may be an option coupled with high speed memory modules. Most likely not cost efficient though, considering EPYC prices.



Your ANSYS license also plays a huge role when determining what hardware you should use.

[Amiga500]

Originally Posted by Simbelmynë

Your ANSYS license also plays a huge role when determining what hardware you should use.

Indeed.

You need to balance the cost of your licenses against the speed of your machine. Is it more cost effective to buy more threads, or is it more cost effective to make those threads work faster.

Going back to when I was last involved with ANSYS - I think a perpetual license was >£50k for 8-thread HPC.

I've no idea what it is today, but its not hard to see that investing in a faster machine than either proposed is likely to pay more dividends rather than buying 16-threads worth of HPC packs and using that Xeon. At work, I have that exact Xeon in a single socket Dell 7910 (other socket on motherboard not used). I think its crap to be honest.


I've a Threadripper 2950 at home and it would be roughly twice the speed per thread (albeit that is on non memory bandwidth limited workloads).

Waiting for the next-gen threadripper and buying a 8 thread license would be more cost effective than running that Xeon on a 16 thread license (and stays with the 2 threads per mem channel rule of thumb).

If you have the money to go above 8 threads, then a single socket EPYC 7371 (or Xeon W-3245) is likely the way forward - with leaning toward the EYPC due to the 8 memory channel config (again sticking to the 2 threads per mem channel rule of thumb). Unfortunately AMD have not yet released a next-gen replacement for the 7371 as part of their recent EYPC 7002 series launch.

If you are going above 16 threads, then your into big money on the license and hardware so should be above to coerce ANSYS and hardware vendors into demos.

[flotus1]

Let's be real...license costs are the dominating factor under normal circumstances. But people asking for help about a decision between a consumer-grade PC and a used mid-range workstation rarely pay full price for licenses. At least I hope not, otherwise priorities would be completely messed up.
Although I have to say I am a bit confused that OP is asking on behalf of a company. Neither do companies get large discounts on licenses, nor is buying used hardware a possibility.

[Amiga500]

Originally Posted by flotus1

Although I have to say I am a bit confused that OP is asking on behalf of a company.

That's why I assumed that it was legit licenses.



I suppose the way the question in phrased brings me to another point I should have thought of earlier. The original poster's company may get better returns by subcontracting out the work to specialists or getting at least one in as a subcontractor for a few months at the start to get things up and running in terms of best practices etc.

As you of course know, but the OP may not, if you don't get the foundations right (suitable mesh and appropriate model which is correctly setup and your approach validated as best as possible), then CFD really does stand for Colourful F**king Diagrams and little else as it almost certainly won't tell you much of use.

[flotus1]

Reading the initial post more carefully, I conclude that we missed an important detail: both machines are already on site.
So yeah, buying Ansys licenses for tens of thousands of dollars per year, but cheaping out on the hardware that can be used for 3-5 years, would not make any sense.
But I could imagine it happening in some companies.

tl;dr: buy a new, proper workstation in the 10k-15k price range.

[vek123]

Originally Posted by Kevin.B

The company I work for is weighing their options for when they buy Ansys to do turbomachinery design. We are hoping to get as many turbine calculations in a day as possible. We have two computers that we think might work, but do not know which one to use, or if we need to build a much more powerful workstation.

Option 1:
i9-9900k
4x16gb dimms of DDR4-2666
Quadro P1000
Windows 10


Option 2:
Dual socket E5-2630v3
8x8gb dimms of DDR4-2133 ECC
Quadro P1000
Windows 10


If neither one of these will be fast enough, would an AMD Threadripper or Epyc based system be better? Or a newer Xeon base system.

You should first decide whether you need high clock speed ( fast for highly serial problems small mesh and small time step dependent ), or higher number of cores ( fast for large mesh problems ).
sometimes a high clock speed core i-5 can be much faster than a multi-core workstation with expensive xeon cpu with lower clock speed.

[Duke711]

Fluent not scale with cpu frequency. Only with memory bandwith and number of cores by a peak of two cores per memory channel.