PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : General discussion : Run Time VS CPU Time? (Message 137586)
Posted 348 days ago by allenout
What is the difference between Run Time and CPU Time?
2) Message boards : Number crunching : Intel or AMD for LLR Tasks? (Message 137585)
Posted 348 days ago by allenout
Intel or AMD for LLR Tasks?
Broadly speaking has one got an advantage over the other at PrimeGrid?
I'm doing a new build and deciding which way to go.

Simply looking for whatever CPU is fastest at LLR...

Is one better than the other for long tasks like SOB / WOO / CUL, or for short tasks like PPS / SGS?

I've been looking at the i9-9900KS / Ryzen 9-3950X but also the i9-10980XE.

Now my son (teenager) is telling me I have to get a Ryzen Threadripper!

Does Primegrid have a table or chart showing the fastest CPU's for a subproject?
Say if we pick PPS MEGA, can we see a 9900KS did it in 800 sec, a 9900 = 900 sec, a 9700 = 1000 sec etc...

Thanks!

If you can do it, intel 99xx-X are the best.
X chips have dual AVX-512 (double check that they do before buying)
Also the available cache can be considered to be L2 + L3 on X chips
They however probably use a lot of power

Threadrippers aren't very good at LLR

Latest from AMD are equivalent to intel non-AVX 512 chips
However their L3 cache is split. So consider these AMD chips as actually 2 chips. You don't want to do a task muti-threading using all cores.

I think I haven't explained it very well, but maybe it is a start?


With Threadripper/AMD you should always single thread then?
3) Message boards : Generalized Fermat Prime Search : GFN-16 MEGA Prime Search starts March, 14, 2019 (Message 137113)
Posted 354 days ago by allenout
[...]If this follows other LLR work, I'd expect optimal running configuration to be 2 cores per task for pretty much all recent CPUs.[...]

I turned off HT on my 3970X, so it's running 36 threads on 36 cores. Then I ran an experiment, running just a single task:

8 cores: 1:41
4 cores: 1:29
2 cores: 1:55

Those time are really good, and it seems that 2 cores is the best choice on this machine for maximum work. But, once I cranked up the number of tasks to use all the cores, 2 per task, times dropped back to the 4-6 hour range. Not really sure why this would be, considering it has 128mb of L3?



I noticed the same thing happening to me with my 3900x. Running 1 task with 3 threads is much faster than running 4 tasks with 3 threads each. I'm guessing windows isn't smart enough to keep each task confined to a single CCX and it's wasting a lot of time getting data from cache in other CCXs. The core clock is not much different for me in either scenario.


Maybe you should try Linux.
4) Message boards : Number crunching : Google Colab and its free GPUs (Message 133529)
Posted 474 days ago by allenout
This keeps repeatedly coming up:
Error obtaining number of platforms (clGetPlatformIDs/1)
Error code -1001, message: Unknown
No compatible OpenCL devices found
5) Message boards : Number crunching : Google Colab and its free GPUs (Message 133527)
Posted 474 days ago by allenout
1. Have a folder called "gfn_sieve" in your Google Drive with "gfnsvocl_linux_x86_64" in it.


How do you get "gfnsvocl_linux_x86_64" in it?
6) Message boards : AP26 - AP27 Search : AP27 Closing??? (Message 133338)
Posted 482 days ago by allenout
to eventually update WWWW and move it to BOINC.

WWWW?
7) Message boards : General discussion : Potential for cheap supercomputer array for PrimeGrid (Message 133042)
Posted 489 days ago by allenout
In August AMD launched EPYC Rome based on their Zen 2 technology and chiplet design.
This has taken the Supercomputer and HPC world by storm due the efficiency and performance of the processors.
The 64 core processors can get over 2 Tflops of performance alone.

My proposal here is that rather than just the community provide their free computing performance to make PrimeGrid work, users could donate a small amount of money to build and maintain essentially a supercomputer grid. This would be great for people who want to help but can't directly volunteer their machines and would rather donate directly.

I worked out that if everyone who uses PrimeGrid donates just $15 this year or $1.25 per month, then we could effectively double the throughput of PrimeGrid from 2 Pflops to 4 Pflops. Task which have previously years to complete could take a year or less.

The EPYC processors I would think are the most appropriate for the task
EPYC 7402P 24C/48T 2.8/3.35Ghz L2:24x512KiB L3:128MiB 180W $1250
EPYC 7502P 32C/64T 2.5/3.35Ghz L2:32x512KiB L3:128MiB 180W $2300
EPYC 7702P 64C/128T 2.0/3.35Ghz L2:64x512KiB L3:256MiB 200W $4425
(Add around $700 for motherboard and RAM per system)

Just throwing the idea out there.
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2021 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 3.56, 3.70, 3.87
Generated 22 Jan 2021 | 13:38:14 UTC