Hello,

I would like to build a dedicated CPDN rig (i.e. will run CPDN only, 24/7).

I tried to search for some configuration guides on the forum but the information I found seems to be a bit outdated, so I decided to open up this post to collect some advice.

For me, the budget is not too much of a concern, so I'm basically looking for the most powerful configuration (within a relatively reasonable consumer budget of course, let's say <= USD 2.5k?)

However, I do face a constraint that I will only be allowed to build one physical machine. So although I do understand that in terms of total throughput it may sometimes be better to build two cheaper machines rather than a single expensive one, in my case that is unfortunately not an option of mine.

That said, I'm ok with those server boards that allows for multiple CPUs on one board. Again, as long as it fits into one physical machine I'm good. So basically all I can do is to maximize the total throughput out of a single machine.

From the top host ranking it seems that CPU core count is kinda more important than single thread performance and AMD is not doing badly at all on CPDN?

So in that case, would Ryzen 2990WX be kinda the most powerful consumer-grade CPU for CPDN? But how would it compare to Intel Core i9-7980XE?

And how about comparing them to two server-grade ES CPUs like two EPYC 7551 or two Intel Xeon E5-2699 v4?

I'm trying to make such comparisons as on eBay these configurations seem to have similar prices:

2990WX ~ US $1,704.99
i9-7980XE ~ US $1,600.00
2x EPYC 7551 ~ 2x US $699.00
2x E5-2699 v4 ~ 2x US $846.00

Does anyone have experience with the ES (engineering sample) CPUs from eBay? Do they perform well and are they reliable enough?

Also, in general, may I ask how I should estimate the performance of CPUs on CPDN? What should I look for? Is there a CPU benchmark score similar to that from PassMark that can act as a reasonable estimator for CPDN performances?

Meanwhile, I noticed that memory bandwidth is another thing that is commonly brought up as a performance limiting factor. May I ask how I should take that into account when it comes to CPU selections? Would Ryzen 2950X or other Intel Core ones with higher single thread performance give more total throughput by any chance because of that? What is the threshold that I should start to be concerned with memory bandwidth?

Sorry if I'm asking some very silly questions. I should have done more research to make up for my lack of computer hardware knowledge, but unfortunately I am not able to find enough time for that in the coming weeks and I don't think I can wait until then to start building the rig.

So I would really appreciate some help.

Any advice is warmly welcomed!

Thanks a lot in advance!!!

A few things to keep in mind:

1) There are usually only a few hundred to a couple of thousand models released at a time, and there's already thousands of processors waiting to get models.
So a huge cruncher may sit (mostly) idle for weeks.

2) It's recommended that 1.5 to 2 gigs of ram per core is available.

3) It's mostly Windows versions of the models that are used these days.

4) Linux models need 32 bit libraries.

5) This isn't a set and forget project. You need to keep an eye on each of your models to make sure that nothing has started going wrong.

I have been running eight wah2 tasks on my new Windows 10 PC with a Ryzen 5 1400 CPU. Of these 7 have failed but 5 have given me credits all the same.I am running the eighth.
Tullio

---

Larfy Hu -

You titled this thread "Most Powerful CPU". You probably mean "Most Powerful System" which involves several components - CPU speed (GHz), maximum memory chip transfer rate, bus speed, amount of CPU cache, and maybe others.

What is best for CPDN may or may not be the best for other work (projects).

The word "Most" implies you want to build the world's fastest system for CPDN. I doubt the information exists on how to do that. If a F1 (auto racing) analogy makes any sense to you, you are trying to build a Mercedes or a Ferrari. But even the worst car in F1 is still a really, really fast car.

Go with the fastest memory you can find and a CPU and main board that supports that speed. All other things being equal, go with the largest CPU memory cache.

As Les mentioned, LINUX work is unavailable because most tasks fail. The people in charge are either unable or unwilling to fix this issue.

CPDN work availability is generally sporadic. So you need a backup plan such as World Community Grid (WCG) or your computer will be idle for long periods of time.

CPDN work does not use Graphic Processors (GPU) and neither does WCG. WCG has several Climate Science projects waiting in the wings, if Climate Science is your main or only interest.

If Climate Science is your only interest, I know of no plans anywhere for climate work that uses a GPU, so you don't need the biggest and best graphic card.

These are my thoughts. If others disagree, please comment.

---

The only thing I would add is that if going for a multi chip system, make sure you have a fast hard disk. The new global models produce over 1GB/day even with only 2 cores active on my laptop which is slow. A fast machine could with these tasks be producing well over 10GB/day and that would just be on 4 cores. If your connection is more bored than broad, don't even think about running those ones!

Edit:Over the years there have been a number of multicore machines dedicated to running cpdn (32 plus cores) that have had very high failure rates and one suspected reason for this is not being able to keep up with the i/o

My newest PC, a HP Omen, has a 128 GB SSD and a 1 TB hard disk.The SSD is not a 2.5" disk like a laptop but a flash memory bolted to the main board.

Tullio

---

I have found to aim for a "slower" front side bus speed, middle of the option available on a motherboard. exit and shutdown BOINC before a PC restart.
Most failure are from PC re-starts, and too fast settings on the RAM, from my findings.
If I was building a PC now I would buy the AMD threadripper and lots of RAM.
Don't overclock.

---

Most failure are from PC re-starts, and too fast settings on the RAM, from my findings.
If I was building a PC now I would buy the AMD threadripper and lots of RAM.
Don't overclock.

Excellent advice. If I had a Threadripper, I think I would start out using full cores by disabling SMT in the motherboard BIOS. That way, you don't overload the memory size or speed requirements. Then, I would enable SMT and start increasing the number of cores to see what effect it has on the speed. It might affect the error rate also.

And don't forget that climate modelling relies heavily on floating point maths.
Back near the start, 80 bit floating point was used. I don't know about now.

So there's probably a heavy demand for the fpu as well as well as custom code, and I think that the Intel processors had the edge there.

A wah2 task ended in 11 days 16 hours and 27 minutes on my Ryzen 5 1400 CPU.
Tullio

---

A wah2 task ended in 11 days 16 hours and 27 minutes on my Ryzen 5 1400 CPU.

Doesn't mean a lot without knowing the region, resolution and number of months the task was for. On my admittedly slow laptop, WAH2 tasks on it at the moment vary between about 3 and 8% a day.

Global, one month, 12 years
Tullio

---

My EPYC 7401p rig is current No.1 with over 200k average and this is still going up.
Think that the 8 channel ram is helping.
I'm looking to put a duel EPYC rig together with some sample chips but you need to be careful with the bios versions for the 7551 samples.

Think that the 8 channel ram is helping.

Something like that will be the wave of the future. The memory speeds just can't keep up with the new CPUs.

The memory speeds just can't keep up with the new CPUs.

Or even the old ones sometimes. I have a laptop with a Pentium N3540 @ 2.16GHz and a short one month task took over 18 hours something rather than 2 hours less when running alongside 3 global tasks. On this machine running a fourth global task slows the others down by almost 25% so virtually no gain from running four at once.