Now that would be a beauty to own.
I love that AMD are putting the stones to Intel.

But one would probably need a minimum of 256GB of memory to keep it responsive and busy but if you were to run 128 FAHB (one per thread) on that machine you might spill over to the swap file occasionally. I know, It happened to me..

Primegrid is the only thing with avx512 support so far as I know.
I think a lot of it comes down to project funding and trying to keep everything open to all types of processors. I'm sure a dual socket Sandy bridge Xeon with 32 threads will out perform my Ryzen 1700x and 1800x in points per day on this project simply because there are 32 WUs being crunched simultaneously. A project like CPDN would see drastically different results per Workunit as it seems to require different parts of the processor and that only real cores be used for speediest results. The Ryzen might be twice as quick per WU at this point, though I can't be sure, whereas on WCG with MCM I might only see a 20-40% difference, if even that, which would then be negated by twice as many WUs on the Xeon. Still, this 128 thread Ryzen would run circles around anything at this point.
WCG isn't memory or hard disk hungry like some other projects can be (CPDN, Rosetta) unless we're talking about arp or even fight aids at home...I think they average around 400 MB per WU?
A lot of gpu projects have CPU to go along with them because there are things gpus just can't crunch. I only have an rx570 so I'm glad there are CPU only projects, since that's what I own most of, and a lot of GPUs have done away with higher double precision computing, if I remember correctly. AMD still remains in the lead there - my rx570 handily out performs a gtx 1080 on that project. Maybe it's FP 64 I'm thinking of - I honestly can't remember specifics.

I'm running a 128 thread EPYC server with 256GB of memory and I occasionally have data on the swap file. Most of the time it is zero but there are occasions when it isn't. That server is 100% dedicated to WCG and has no desktop code installed on it and running 19.10 of Ubuntu Linux. I've noticed an interesting anomaly on systems with large memory configurations (above 32GB) related to Rosetta and Gromacs mostly. If I run Rosetta (MIP) on systems with 32GB or smaller the memory allocation is mostly as you describe. Once you get above 32GB these programs seem to allocate more memory per process. It is really noticeable on the virtual allocation but also on the real side but to a lesser degree. Rosetta allocates 12.2GB virtual per process and about 1.3 or so real on the 256GB system. On the systems closer to 32GB they allocate about 3 GB virtual and then under 1GB real. On 64GB systems, it falls in between the two like 6GB or 7GB per process and about 1GB real. I captured memory debug information and sent it to WCG at their support email address. Their response was, they don't test on systems that large and so they wouldn't see it. They are forwarding the info to the developers. Anyway, since the real allocation is still reasonable it doesn't create a big impact and I can run 128 FAHB simultaneously. But when you start running 32, 64, 128 or 256 threads, the amount of storage allocated to buffers, cache, and other things is almost as much as the programs themselves. Right now on the 128 thread system with mostly MCM1 (120) and 8 CPDN (8) units. Memory utilization is 21GB for processes but over 31GB is allocated to buffers and cache. The anomaly doesn't happen on MCM1 mostly Rosetta. I tried running 128 Rosetta with 128GB and it didn't work. To balance out the memory across the memory channels, I went ahead and ordered another 128GB and it's been fine. I have to be careful what projects I run as CPDN and other projects together are too large even for 256GB.

I'm more curious what kind of cooling can keep this beast under control crunching 24/7. That's my only concern right now. I have all the gears ready for it, just waiting for the cpu itself and the cpu cooler. If I can't find a good cooler then I may have to go with its lil brother, the 3970X. The Kraken X72, deepcool castle 360ex are the top AIO coolers and they can barely hold the 3970X on light weight projects. Run MCM full load, temps will be in the high 70s/low 80s. AVX? good luck with that :D