Here's the deal: You get more price with more performance.

The best of AMD comes from the 3950X, but threadrippers really aren't that good at LLR. They're priced at about $699. They don't have AVX-512 and are about equal or higher to i9-9900K. They have MUCH more cache than 10980XE, about three times.

The best of Intel HEDT comes from 10980XE, excluding Xeon Platnums which aren't HEDT. They HAVE two AVX-512 units and benefit largely from the LLR V3.8.23. They are higher priced at $999.

So it depends on how much money you have and how much you want to spend on your GPU.
I prefer 3950X though its cache is split among CCX's. It's ultimately just whether you like intel or AMD more ;)
____________
My lucky number is 6219*2^3374198+1

I am seeing a bunch of conflicting and wrong info here.

Anything Zen/Zen+ crippled AVX2, so avoid that. And comparisons to those obsolete chips are not relevant.

Zen2 has full AVX2/FMA3. Same as normal Intel chips.

WRT Intel chips with AVX-512, those chips are faster per LLR task...sorta. It really depends on the number of cores, cache size, the number of threads assigned per task, etc. Also the choice of LLR sub-project can make a larger difference, coupled with the choice of number of threads.

Finally the amount of cores with CPUs with AVX-512 are relatively small compared to the threadripprer 3xxx. For example a 32 core threadripper with AVX2 may outperform a 16 core intel X with AVX-512.


____________
Reno, NV

I have the same problem here, and its been continuing for about since I had this router...
____________
My lucky number is 6219*2^3374198+1

Zombie's reply is best so far.


Let's look at it from an execution potential perspective. Assume there are no cache or ram limits, how good is the architecture if you compare say 1 core at the same clock?

Take Intel consumer mainstream desktop CPUs from Skylake onwards as the reference point. Zen and Zen+ are about half the performance and basically shouldn't be a consideration unless really cheap and you don't care about density. Zen 2 (used in desktop Ryzen 3500 or higher) catches up and even slightly exceeds Intel, although I never worked out exactly what that number is. Consider it as "slightly more" for now. Two unit AVX-512, which includes Skylake-X and Cascade-X HEDT CPUs, and some Xeons, have about double the potential of the mainstream Intel CPUs.

So for rough indication of peak CPU potential, I'd literally take the cores and multiply by all core loaded clock (note AVX boost is typically lower than for other workloads on most CPUs), and for the AVX-512 CPUs, multiply by 2. This is a good indication of throughput for smaller tasks like PPS or below.


I said that was assuming no limits, so what are the limits? We have the FFT size used by LLR. The FFT data size is the FFT size multiplied by 8. e.g. 256k FFT would be 2MB of data. Through observation the FFT data size compared to CPU cache is a good indication of performance, even though there is other data used which is not considered. At a simple level, take the number of tasks you want to run at the same time, multiply by the data size, and you have your effective working requirement. If that number is less than the L3 of the CPU, you wont be ram bandwidth impacted, and can expect high performance. If the number is greater than the L3 cache, then you may be limited by ram bandwidth.

There are a number of additional considerations to the above. For Skylake-X, it looks appropriate to include the L2 cache when doing the consideration, but it isn't something I looked at in too much detail. This may be a side effect of those CPUs having a relatively large L2 compared to their L3.

Ryzen CPUs have L3 cache partitioned per CCX, which is 3 or 4 cores depending on the CPU model. Cores within the same CCX can share data easily. If it is not on the local CCX, in effect it has to go back to ram to get data, and you're hit with potential bandwidth limitations there. So for best throughput, it seems best to run Ryzen CPUs as if they were clusters or 3 or 4 cores depending on the model, at least for medium sized tasks. It gets complicated when you get to bigger tasks that no longer fit inside a CCX. There still seems to be some benefit from the large caches, but not as much as if they were unified.

And quickly on multi-threading, the guide I use in the first instance is that best throughput performance seems to be when the caches are nearly filled. So you would select the number of tasks running that would satisfy that. For example, PPSE you would typically run 1 task per core. MEGA is a big bigger and 2 threads per task starts to become useful. Basically, exceeding cache puts you into potential ram limiting, and running more threads than ideal loses you throughput through the lower efficiency scaling as you increase the number of threads per task.


So overall, to answer the original question... It depends on exact pricing to some extent but if Cascade Lake-X is in budget, I'd consider that, or the equivalent Skylake-X if pricing is similar (not higher as it was). These get you AVX-512 and 4 ram channels compared to 2 of the 3950X. The downside is mainly availability, there don't seem to be many of them in stock in general. These can get very power hungry but in part I think it is mobo manufacturers typically defaulting to power unlimited/maximum turbo. Use a more sensible power limit to help tame their AVX-512 consumption. I run my 7920X with turbo off for example.

2nd choice would be the 3950X. It has much more potential over the 9900ks in most things. The Threadripper 3000 series is also interesting and you can consider it as a scaled up 3950X.

With Threadripper/AMD you should always single thread then?

I have a 1st gen ThreadRipper, 1920X with 12/24cores, and in some cases using 11 cores was faster than 23 cores but not for every application here. My older Intel I7-3930K would run the same workunits in about the same amount of time using only 5 or 6 cores, I don't remember exactly anymore. The key for me is what else I do with the pc, if it's only ever going to be here at PrimeGrid imho Intel is faster overall on a per core basis, but if you are going to run other Projects as well then you can get more cores cheaper thru an AMD cpu and do more units in quantity than a similar priced Intel cpu.

No you will not compete with the best from Intel.
AMD throws many cores at computing.
You need to use many cores with threadrippers to be equal with the best from Intel

If you throw all the cores on a 3950X to a fight with a 10920X or a 3990X with a 10980XE of course the one with more cores would win(chinese benchmarks).

With regards to FFT data size and CPU cache, does it matter if cache almost fits FFT data, or is a more of a simple, yes it fits or no it doesn't determination?

For example, if FFT data size is 10 and cache is 8, is that better than if cache is 2, or is it equivalent since in both cases FFT data size is bigger than cache?