[artymk4]

I am looking to buy a PC for the next 6 years. I will be doing CFD but also other things. Here are answers to the checklist:
I will be using Ansys Fluent
I have a license for 12 cores but sometimes I am limited to 4 or 8 cores so 4 / 8-core performance is also important.
Mostly steady CFD simulations, occasionally transient. Mostly small meshes up to 2M cells, occasionally up to 30M. I do a lot of optimizations with 100s of design points.
Budget is roughly 2000 EUR.
Used parts are not an option. Can be OEM or buying parts. I'm from Europe.

I did some research but I can't decide between a desktop PC and a server.
For a desktop I am looking at i9-12900KS or i7-12700K with 64 Gb DDR5 @ 6000 MT/s. There are no dual CPU motherboards available with socket LGA1700 right?
For a server I am looking at AMD EPYC 7313P with 8x 8 Gb DDR4 for example: link
Or would you suggest something entirely different?

[the_phew]

You'd typically need ~80+ GB/s of memory bandwidth to feed 12 cores with a CFD workload, and/or a CPU with a ton of cache. You also probably need at least 128 GB of RAM if you are running simulations with 30 million cells regularly, but I've never used Fluent so I'm no authority on that matter. That budget likely rules out current-gen server CPUs, but server CPUs may not offer you much with only 12 threads anyway. You might be able to swing an EPYC 7272, but a 7313P would probably blow your budget out of the water (especially with 128 GB of RAM).

Alder Lake is enticing, but DDR5 is expensive, those CPUs have weird memory controllers and seemingly can't take advantage of memory-bandwidth intensive workloads when populated with DDR5, you're practically limited to 64 GB of DDR5 (both by budget and lack of memory/motherboard options that support 128 GB right now), and you get at most 8 'performance' cores to run on anyway (running on the 'efficiency' cores would likely just slow you down, unless ANSYS was very proactive about software updates for Alder Lake).

The Ryzen 5800X3D only has two channels of DDR4, but all that L3 cache helps a LOT for CFD, especially for smaller simulations. If you can wait a month or so, AMD should be releasing the 5900X3D, which should be nearly perfect for you; 12 cores and a whopping 200 MB of L2+L3 cache.

If you can't wait, Intel's venerable i9-10920x is still a viable option in your price range; it should be at least as fast for CFD as any currently-available desktop processor, thanks to its four memory channels. It just feels dumb to buy an (effectively) 5 year old processor new.

[nmc1988]

Originally Posted by the_phew

You'd typically need ~80+ GB/s of memory bandwidth to feed 12 cores with a CFD workload, and/or a CPU with a ton of cache. You also probably need at least 128 GB of RAM if you are running simulations with 30 million cells regularly, but I've never used Fluent so I'm no authority on that matter. That budget likely rules out current-gen server CPUs, but server CPUs may not offer you much with only 12 threads anyway. You might be able to swing an EPYC 7272, but a 7313P would probably blow your budget out of the water (especially with 128 GB of RAM).

Alder Lake is enticing, but DDR5 is expensive, those CPUs have weird memory controllers and seemingly can't take advantage of memory-bandwidth intensive workloads when populated with DDR5, you're practically limited to 64 GB of DDR5 (both by budget and lack of memory/motherboard options that support 128 GB right now), and you get at most 8 'performance' cores to run on anyway (running on the 'efficiency' cores would likely just slow you down, unless ANSYS was very proactive about software updates for Alder Lake).

The Ryzen 5800X3D only has two channels of DDR4, but all that L3 cache helps a LOT for CFD, especially for smaller simulations. If you can wait a month or so, AMD should be releasing the 5900X3D, which should be nearly perfect for you; 12 cores and a whopping 200 MB of L2+L3 cache.

If you can't wait, Intel's venerable i9-10920x is still a viable option in your price range; it should be at least as fast for CFD as any currently-available desktop processor, thanks to its four memory channels. It just feels dumb to buy an (effectively) 5 year old processor new.

Based on the result of Openfoam benchmark
i9-10920x is slower than i5-12600 (16 cores compared to 6 cores respectively)

OpenFOAM benchmarks on various hardware

OpenFOAM benchmarks on various hardware

In total Dkr 7.533 which is € 1005 or $ 1147 (prices are excl V.A.T)
The components are:

Intel i5-12600 (Dkr 1599,20)
Kingston Fury Beast 2x16 GB 6000 MHz (Dkr 2636,80)
Asus TUF Gaming B660M-plus wifi (Dkr 1247,20)
Samsung 970 EVO plus 1TB (Dkr 709,60)
Noctua cpu (NH-U9B SE2) and cabinet coolers (Dkr 525,60)
Seasonic Focus GX550 (Dkr 479,20)
Fractal Design Core 1100 (Dkr 335,19

[the_phew]

Originally Posted by nmc1988

Based on the result of Openfoam benchmark
i9-10920x is slower than i5-12600 (16 cores compared to 6 cores respectively)

OpenBenchmarking.org has the i9-7960X trading blows with Alder Lake and Ryzen 5000 in OpenFoam Motorbike 30m, which is remarkable for a 5 year old chip (that Intel still sells as a current product in the same price range as those desktop chips as the i9-10920x et al):
https://openbenchmarking.org/test/pt...f51edd#metrics

I still wouldn't recommend those 'Extreme edition' processors for a new build, since they are ancient tech now. But it just drives home the point that 4+ memory channels go a LONG way for CFD when you are talking 12ish threads. Especially since Fluent isn't OpenFOAM, and for any generic CFD solver, "When in doubt, go for memory bandwidth". But extra cache seems to provide 'effective' memory bandwidth for CFD as well, which is why my ultimate recommendation is the forthcoming 5900X3D if the OP can wait a month or so.

The combination of the hybrid architecture providing ample opportunity for CFD solvers to run very slowly (if process affinity isn't correctly set up either in the OS or the solver), and well-documented issues with it's memory controller not exploiting DDR5's memory bandwidth makes Alder Lake a risky recommendation.

[wkernkamp]

Originally Posted by the_phew

......

The combination of the hybrid architecture providing ample opportunity for CFD solvers to run very slowly (if process affinity isn't correctly set up either in the OS or the solver), and well-documented issues with it's memory controller not exploiting DDR5's memory bandwidth makes Alder Lake a risky recommendation.

But he did prove performance with actual hardware?

[artymk4]

Thank you everyone for the help. I decided to wait for Ryzen 5900X3D however it's still not certain whether they will release it or not. Ryzen 7000 series is around the corner (September 15th).

Nobody commented on the OEM workstation options. I posted one link in first post. It features an Epyc 4313P which is 8 memory channel CPU but the motherboard has only 6 DDR4 slots. How much of a bottleneck that is?

I found another offer from Lenovo with big discount so it almost fits my budget:
https://www.lenovo.com/us/en/p/works...620/30e000tcus
It has Threadripper PRO 5945WX (12 Core) which is also 8 memory channel. The workstation includes 2x 16Gb RAM but I assume another 2x 16Gb can be added (would this void the warranty?).

My guess is that 4-core performance of Ryzen 5900X3D will be as good if not better then the above workstations, however the workstations are better in 8 or 12-core performance. Do you agree?

[the_phew]

If you can swing a Threadripper Pro 5945WX within your budget, that is probably the fastest/least risky 12 core option for CFD. Lenovo wouldn't void your warranty for adding RAM, just make sure you get the EXACT same model RAM sticks as what it comes with. 128 GB is really the least memory you can get by with if you are running larger CFD simulations, so you'd want to add more 16GB sticks anyway.

The upcoming 5900x3D will likely be faster for smaller grids (it has half the memory bandwidth of the 5945wx, but 3x the L3 cache). But the Threadripper Pro will be much more versatile, with much higher max RAM capacity, all that memory bandwidth, MANY PCIe lanes for GPU compute down the road if you ever need that capability, etc. And the 5945wx is available now, while the 5900x3D hadn't even been released.

[wkernkamp]

Originally Posted by artymk4

Nobody commented on the OEM workstation options. I posted one link in first post. It features an Epyc 4313P which is 8 memory channel CPU but the motherboard has only 6 DDR4 slots. How much of a bottleneck that is?

I found another offer from Lenovo with big discount so it almost fits my budget:
https://www.lenovo.com/us/en/p/works...620/30e000tcus
It has Threadripper PRO 5945WX (12 Core) which is also 8 memory channel. The workstation includes 2x 16Gb RAM but I assume another 2x 16Gb can be added (would this void the warranty?).

My guess is that 4-core performance of Ryzen 5900X3D will be as good if not better then the above workstations, however the workstations are better in 8 or 12-core performance. Do you agree?

Odd that an Epyc 4313P board would have only 6 DIMM slots. Probably the board does have eight and not six.


The Threadripper PRO 5945WX workstation looks like a good deal to me. From the pictures, you can see that it can carry eight dimms (compatible with the 8 channels). For best performance all channels should be active so the price will go up.


I agree that for 4-core performance the Ryzen 5900X3D will be hard to beat. It has an enormous cache and two ver fast memory channels (which is enough for 4 cores.)

[flotus1]

Threadripper Pro 5945WX is not a good fit for a CFD workstation.
It technically has 8 memory channels. But 64MB of L3 cache tells us that only 2 out of 8 chiplets are enabled. Which only provides enough bandwidth equivalent to ~4 memory channels. Not that 12 cores would benefit massively from 8 "real" memory channels. This CPU just isn't good value no matter how you look at it.
Further information about this topic: https://www.servethehome.com/amd-epy...ory-bandwidth/

[the_phew]

Even with only four memory channels, the 5945WX still surely has enough memory bandwidth to feed 12 cores. There really aren't any other good options that have 12 high-clocked cores and 4 memory channels, besides the aforementioned i9-10920X (which can be readily overclocked to 4.4 GHz all-core), which is too ancient to recommend.

Having just built a 5950x workstation for occasional CFD that cost us over $2k USD (128 GB RAM), there really isn't enough memory bandwidth to scale linearly above 8 cores. I'm also regularly using a 3955WX (with the same neutering to four memory channels as the 5945WX), and it scales linearly up to all 16 cores. So you really need the following to squeeze the most out of 12 cores for CFD:
-Fast CPU clocks
-Four memory channels of DDR4 OR two channels of DDR5 OR two channels of DDR4 plus gobs of L3 cache

So the 5900X3D is perfect, but not released. The regular 5900X will be starved for memory bandwidth with larger grids. The 7900X MIGHT be less than a month away, so it's another good option (DDR5 for gobs of memory bandwidth, insane clocks, without the hassles of Intel's hybrid architecture). A 12700K w/DDR5 would MAYBE work well, but the hybrid architecture can ruin parallel performance in solvers not optimized for it. The 5945WX would surely peg all 12 cores all the time for almost any reasonable cell count, it's just overpriced for what it is. But if the OP can score a deal on a workstation with that chip (then add more matching RAM purchased cheap from a third party), there is no risk of it not meeting expectations.

If the OP can wait 6 weeks or so, there should be better options than now. The Threadripper Pro 5000 chips will only get cheaper in the meantime, too.

[martinh]

Originally Posted by the_phew

Even with only four memory channels, the 5945WX still surely has enough memory bandwidth to feed 12 cores. There really aren't any other good options that have 12 high-clocked cores and 4 memory channels, besides the aforementioned i9-10920X (which can be readily overclocked to 4.4 GHz all-core), which is too ancient to recommend.

Having just built a 5950x workstation for occasional CFD that cost us over $2k USD (128 GB RAM), there really isn't enough memory bandwidth to scale linearly above 8 cores. I'm also regularly using a 3955WX (with the same neutering to four memory channels as the 5945WX), and it scales linearly up to all 16 cores. So you really need the following to squeeze the most out of 12 cores for CFD:
-Fast CPU clocks
-Four memory channels of DDR4 OR two channels of DDR5 OR two channels of DDR4 plus gobs of L3 cache

So the 5900X3D is perfect, but not released. The regular 5900X will be starved for memory bandwidth with larger grids. The 7900X MIGHT be less than a month away, so it's another good option (DDR5 for gobs of memory bandwidth, insane clocks, without the hassles of Intel's hybrid architecture). A 12700K w/DDR5 would MAYBE work well, but the hybrid architecture can ruin parallel performance in solvers not optimized for it. The 5945WX would surely peg all 12 cores all the time for almost any reasonable cell count, it's just overpriced for what it is. But if the OP can score a deal on a workstation with that chip (then add more matching RAM purchased cheap from a third party), there is no risk of it not meeting expectations.

If the OP can wait 6 weeks or so, there should be better options than now. The Threadripper Pro 5000 chips will only get cheaper in the meantime, too.

This was really interesting to read, thank you.

Do you have insights on the newly released Ryzen 7950x? It has larger caches but at the same time with 128GB of RAM the base clock speed goes down to 3600 MT/s (I've seen some posts reaching 4200 after lots of struggling...) given that it has only 2 memory channels and you'd need 4 sticks of 32GB currently.

Would this be a better option nowadays than the prohibitively expensive Threadrippers such as the 5965WX which I had been eyeing but cannot afford?

[the_phew]

Originally Posted by martinh

Do you have insights on the newly released Ryzen 7950x?

Due to the memory limitations you mentioned (dropping down to 3600 MT/s when using two DIMMs per channel), the 7950X is only marginally faster than the 5950X for CFD if you need 128 GB of RAM, at least until 64 GB DDR5 desktop DIMMs arrive (AND get supported by motherboard manufacturers).

We're currently in a market funk for CFD on sub-$1000 desktop CPUs. There really hasn't been a better option than the i9-10980XE in the three years since its release. The Threadripper Pros and Milan-X CPUs are CFD monsters, but a huge step up in price.

I'm hoping the Ryzen 7950X3D will finally be the sub-$1k CFD CPU to get when it's released in a few months. But just like the 7950X, it'll be neutered until 64 GB DDR5 desktop DIMMs are released and supported.