Any quick thoughts from you who did a LOT of research
 

Looking to build a new PC .... I think I can build 1 that can match the thruput of my 4 in the dungeon for less power.
There are 3 i5-3570 and 1 i5-2500

My i7-7820X with 32GB RAM (24 allocated to GIMPS) does more than the 4 combined.

There are a lot of new PCs chips lately ... and yeah I could and if necessary I will do the necessary comparisons.
But just in case someone already did that I'd appreciate any thoughts.

I see new PCs anywhere from $500 to $3500 and based on a quick benchmarks scan spending a LOT more money does not seem to add that much more thruput.

I suspect I agree with others that I want a PC that has a near top performance with the least $$$ and the least KwH.

Thanks in advance.

PS I've always had intel but am NOT opposed to AMD.
PPS I'll probably also add a GPU; NVIDIA 30xx.

For small FFTs, I would recommend some Intel chip with AVX-512 like the newer X series or the 11xxx series. The latter only goes up to eight cores and I would recommend against the 11900(...), instead I would recommend the 11700K/KF. An additional possibility would be 12xxx with efficiency cores disabled and AVX-512 enabled, but unfortunately, this is not possible with the mayority of new chips anymore.

For "bigger" FFTs, I would recommend a AMD Zen 3 chip, e. g. 5950X. Spending more will lead to Threadripper, but also a lot more memory throughput.

If you want me to run some benchmarks, I am glad to help you.

I ordered a 5950X Monday. Seems the best all-round chip, but I don't run P95 primality tests so memory bandwidth is not the top priority for me. For the workflow I have (GMP-ECM, nfs factoring, LLR on 2M- to 5M-bit numbers, some P-1 for your project), the 5950 was the best choice.

If huge P-1 is one of your objectives, try to find a platform (mobo/cpu combo) that supports huge amount of RAM. Most of them should support 4 sticks of 32GB RAM, so 128 GB is your baseline. See if you can find something that supports 256 (or higher!).

The Ryzen 9 5950X seems like a great choice, it recently had a decent price drop here. It supports up to 128GB of RAM (unofficially even ECC RAM), extreme performance and good power efficiency.

The chiplet design can pose some problems though. On very small exponents V30.8 stage two can choke if using all cores. With a 5900X I had to set up two workers (one per chiplet), run stage 2 on one worker and set the other chiplet as Polymult helper threads in prime.txt. If I run just one worker with all cores, the performance actually tanks.

For bigger FFTs the 5900/5950 might get blown out by the 5800x3d thanks to 96MiB of L3 cache. If DDR4 bandwidth is no longer the main bottleneck (which remains to be seen) it may not be necessary to go full dual channel dual rank 3200+ which could save some pennies and power.



If the workload is wavefront PRP you're probably better off adding a Radeon VII or RDNA2 to the cheapest efficient CPU you can find (maybe a Ryzen 5500/5600?) if GPU prices normalise, nvidia is best for TF not PRP but my info may be outdated. Doubtful intel dGPU's are in the picture first gen, maybe next year.

[QUOTE=VBCurtis;603439]I ordered a 5950X Monday. Seems the best all-round chip, but I don't run P95 primality tests so memory bandwidth is not the top priority for me. For the workflow I have (GMP-ECM, nfs factoring, LLR on 2M- to 5M-bit numbers, some P-1 for your project), the 5950 was the best choice.[/QUOTE]

Love my 5950x, getting about 2100s a test for 3m bit ranges running 16 LLR instances.

Thanks for the comments so far...
 

BTW my current plans are to focus on low and deep P-1 (under 20M).

I get the sense the majority are recommending AMD 58xx/59xx.
Does my work preference change your mind or support it?

I can get either of these for about $850 CAN.
Core i9-12900K: 8 P-Core & 8 E-Core... so not a real 16 core?
Ryzen 9 5950X: Real 16 core?

Similar clock speeds and turbo and RAM support.
AMD less power.

Intel: (Specs not listed for AMD)
Memory Channel: 2
Max Memory Size: 128 GB
Max Memory Bandwidth: 76.8 GB/s

[url]https://www.newegg.ca/Product/Productcompare?CompareItemList=9SIAVH1HR64896%2C9SIADGEGMY1883[/url]

[QUOTE=petrw1;603490]BTW my current plans are to focus on low and deep P-1 (under 20M).

I get the sense the majority are recommending AMD 58xx/59xx.
Does my work preference change your mind or support it?

I can get either of these for about $850 CAN.
Core i9-12900K: 8 P-Core & 8 E-Core... so not a real 16 core?
Ryzen 9 5950X: Real 16 core?

Similar clock speeds and turbo and RAM support.
AMD less power.

Intel: (Specs not listed for AMD)
Memory Channel: 2
Max Memory Size: 128 GB
Max Memory Bandwidth: 76.8 GB/s

[url]https://www.newegg.ca/Product/Productcompare?CompareItemList=9SIAVH1HR64896%2C9SIADGEGMY1883[/url][/QUOTE]


I'm almost wondering if getting an older threadripper model for a similar price for quad channel instead of dual channel would be viable. For example, I'm seeing the 2970wx 24 core system on Amazon right now for ~$800. I'll let someone with more experience chime in.

[QUOTE=petrw1;603490]BTW my current plans are to focus on low and deep P-1 (under 20M).[/QUOTE]On the GPU side, RTX30xx is great for TF and not well suited for P-1 etc because of its relatively low DP performance. Radeon VII or RX 6x00 would be more effective at P-1 than most NVIDIA GPUs, having relatively higher DP performance and good memory bandwidth.

Until such time as the 30.8 enhanced stage 2 is implemented in gpuowl, P-1 on GPU should not be considered. It would be great, however, if gpuowl could be modified to do stage 1 save files compatible with P95. Or if someone could write a converter script ...