Should I get 7950x or wait for 7950X3D
 

I'm sure this topic has been discussed in other threads, but I would appreciate if I get some specific help.
So I have the budget of $2700 which I have to spend on a CPU, MB and RAM.
I was thinking to get i9-10980XE but if I'm going to buy a new setup, I would prefer something with DDR5 support. Threadrippers are out of my reach.
So should I got with 7950x or wait for 7950X3D? Or should I wait, collect some more $ and buy anything else?
Keep in mind, that I also work with Mechanical, not only with Fluent. Usually I need more than 90gb Ram.

Thanks

Ah, the old wait for shiny new hardware dilemma. I'm familiar ;)
The usual solution is simple: if you need it now, buy what's available now. Otherwise wait as long as possible, because there will always be better stuff.
Though this one might be even easier. Zen4 CPUs with 3D Vcache have been hinted at, yes. Though when they will be available, nobody knows. And it is particularly uncertain if there will even be a Ryzen 7950X3D. We had similar rumors for Ryzen 5000, and all we got in the end was the 5800X3D. Though that's not necessarily a bad thing, higher core counts would not necessarily help for CFD and FEA, and only drive up costs.

So you definitely need 128GB of RAM. Would even more help you with your work? Then that could be a solid reason to look for a different platform.

I'd be more inclined to wait for 64GB DDR5 DIMMs (and motherboard support) than the 7950X3D. The Ryzen 7000 will downclock the memory to 3600 MHz (DDR4 speeds) if you populate all 4 memory slots, so until 64GB DDR5 desktop DIMMs are available+supported, the Ryzen 7000 series CPUs have no more memory bandwidth than the Ryzen 5000 series CPUs in workstation applications that require 128GB RAM.

3D cache will be nice, but unless the memory is running at its rated speed, the processors will still be severely kneecapped for CFD/FEA. It's sad that the ancient i9-10980xe still has almost twice the memory bandwidth of the latest desktop CPUs from Intel and AMD when used with 128GB of RAM. For that reason, if you are running CFD/FEA full time, you should probably stay far away from desktop CPUs and stick to server (EPYC/Xeon) or workstation CPUs (i.e. Threadripper Pro or Xeon-W).

Sure, lower memory speeds with lots of DIMMs/ranks are kind of a bummer.
But that's nothing new, all CPUs behave that way. It's just that up until 5 years ago, none of the CPU manufacturers bothered to publish those specs. Not enough people really cared about it for desktop CPUs.
And particularly for this 1st generation of DDR5 platforms, memory support will improve quickly with bios updates. AMDs official spec won't change, but the transfer rates you can actually run with high memory capacity will increase. I highly doubt that DDR5-3600 is the limit right now.

But I agree that this limitation leaves the door open for alternatives. That's why I tried to prod at even more than 128GB, which would make that decision easier.

Yeah, we have a 5950X workstation with 4x32GB DDR4-3600, and it runs the memory at that speed when 'DOCP' is enabled in the BIOS. But that's technically 'overclocking'; it only works because motherboard and memory manufacturers have DDR4 tuning well-honed by now.

I'm sure they'll get there with DDR5, but I don't want to be a guinea pig in the meantime.

So all in all, even based on this topic https://www.cfd-online.com/Forums/ha...dware-wip.html what setup (CPU, MB, Ram) would you recommend me for $3000?
Buying some old 2 socket Xeon is a way to go?

Is there a benchmark site where I can see more detailed results, like this one?
https://www.cfd-online.com/Forums/at...er_scaling.png

I don't know how much core you need but around 2000 dolar you can set 7950x processor 64 GB RAM and 420 mm liquid cooling. This 5.0 GHz 16 cores will definitely compete with old 2 socket Xeon. In addition, you can also use for another areas such as gaming or daily issues if you want to add GPU.

For CFD, a 7950x won't compete with a 40+ core 2-socket Xeon or EPYC node from the past 4 years. You're talking <100 GB/s of memory bandwidth vs. 250-400 GB/s. Unless your CFD solver licensing is capped at 16 cores or scales horribly, the older 2P server node will be at least twice the speed of a 7950x.

A used 2x32 core EPYC Rome server with 256+ GB of RAM goes for like $2k USD and will over 5x faster for OpenFOAM, at least according to https://openbenchmarking.org/test/pt...80ecda#metrics (you have to look at v1.1 of that benchmark to get the 2P Rome results).

Thank you so much for your well grounded answer. I gave my answer according to 16 core limit on some softwares such as StarCCM+. In case of 20-30 million mesh 64 core will give better performance. More RAM capacity is also another good plus.

Which motherboard and cooling would you suggest for 7950x?
Could you provide several budget options?