mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware

Reply
 
Thread Tools
Old 2022-03-21, 02:42   #1
MarkVanCoutren
 
May 2021

5 Posts
Default New CPU questions

Good Evening Gentlemen,

I just upgraded from a 3900X to a 5950X CPU. Based on my speeds and time left, it's about 26.7% faster which is cool but I was expecting it to be close to 33.3% faster just based on the number of cores, not to mention I'm boosting much higher. Is there another bottleneck I could have or is that not how prime95 works? I don't think it's RAM or Power because I have 32 GB and an 850 W power supply.

On a related note, when I was using cinebench23, my temps were a consistent low 60s (C) for the whole 10 minutes, when I turn GIMPS back on to use my idle computing power, it almost immediately jumps to high 60s low 70s (C). I know that's not enough to cause concern but Cinebench always made my old CPU run hotter than GIMPS. Is there a reason? I've got a Noctua NH-D15 with two fans and thermal grizzly kryonaut so there's probably not much better cooling without going to custom water cooling or maybe a high tier AIO. Cinebench uses 100% of my CPU while Prime95 usually hovers around 50%.

I'm on windows 10. 1 worker using all 16 cores. Thank you for any help!
MarkVanCoutren is offline   Reply With Quote
Old 2022-03-21, 03:35   #2
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

2×7×13×29 Posts
Default

Prime95 needs more memory bandwidth than you have available. 32GB ram doesn't matter- RAM speed matters.
Your CPU is fast enough that some cores are waiting on data from RAM, capping the speed of Prime95.

You can test this yourself to some degree by turning off the boost, or turning down the boost- until you reach some slower value, Prime95 will keep reporting the same timings. At some lower speed, memory will finally be fast enough to keep the CPU fed and below that CPU speed P95 will get slower. Or, if you have nice RAM, make sure the BIOS is set to use the higher rated speed of the memory sticks rather than the default speed that "all" memory should work with.

This is why people building Prime95-only boxes buy cheap CPUs but fast memory sticks- it's pretty easy to saturate 2-channel DDR4.
VBCurtis is offline   Reply With Quote
Old 2022-03-21, 03:48   #3
MarkVanCoutren
 
May 2021

5 Posts
Default

Thank you! I hadn't even thought about the speed of the RAM being a bottleneck. Right now I have G.Skill Ripjaws DDR4-3600 CL16 Memory. Is that nice enough that there might be something disabled in the BIOS?

Turns out my RAM was running at like 2133 MHz. I found a different thread that said to check the A-XMP button so I did (profile 2 if that means anything) and it's much faster. I should mention I have an MSI x570 board with the latest BIOS.

Last fiddled with by MarkVanCoutren on 2022-03-21 at 04:01 Reason: Adding a follow up
MarkVanCoutren is offline   Reply With Quote
Old 2022-03-21, 08:07   #4
JWNoctis
 
"J. W."
Aug 2021

1000112 Posts
Default

You might also want to try running 2 workers with 8 cores each. 5950X should still have almost enough L3 cache on each chiplet to run LL double check without hitting RAM. Assignments with larger FFT size might still benefit from better locality, but I myself have nothing with which to try.
JWNoctis is offline   Reply With Quote
Old 2022-03-21, 14:13   #5
petrw1
1976 Toyota Corona years forever!
 
petrw1's Avatar
 
"Wayne"
Nov 2006
Saskatchewan, Canada

3×11×157 Posts
Default

Quote:
Originally Posted by MarkVanCoutren View Post
Thank you! I hadn't even thought about the speed of the RAM being a bottleneck. Right now I have G.Skill Ripjaws DDR4-3600 CL16 Memory. Is that nice enough that there might be something disabled in the BIOS?

Turns out my RAM was running at like 2133 MHz. I found a different thread that said to check the A-XMP button so I did (profile 2 if that means anything) and it's much faster. I should mention I have an MSI x570 board with the latest BIOS.
I got caught in the 2133 vs 3600 scam too.

I have a 3 year old 7820x with 32GB DDR4 3600 Quad Channel RAM and I still get the best thruput with 1 worker 8 cores.
petrw1 is offline   Reply With Quote
Old 2022-03-21, 15:22   #6
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

2×7×13×29 Posts
Default

Quote:
Originally Posted by MarkVanCoutren View Post
Thank you! I hadn't even thought about the speed of the RAM being a bottleneck. Right now I have G.Skill Ripjaws DDR4-3600 CL16 Memory. Is that nice enough that there might be something disabled in the BIOS?
Cool! You just reaped one of the hidden rewards of buying nice kit: a feature you didn't realize was important was included, and you now have a little "free" upgrade! The faster memory speed should slightly speed up regular tasks too, but for Prime95 (and LLR if you search for primes of other forms later) you've seen how that 3600 feature is essential.

You *might* still saturate 3600 speed memory, so you may still benefit from turning boost down a little bit- lower power use and lower temps at "no" cost to P95 speed. I'm considering buying a 5900 or 5950 soon myself, so I am curious if you can quantify your speed improvement from turning XMP on. E.g. what were your old iteration times, and new?

Edit: I am also curious how cinebench and P95 temps are now that memory is feeding the CPU more fully. If you don't mind....

Last fiddled with by VBCurtis on 2022-03-21 at 15:25
VBCurtis is offline   Reply With Quote
Old 2022-03-22, 01:04   #7
sdbardwick
 
sdbardwick's Avatar
 
Aug 2002
North San Diego County

2·3·112 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
Cool! You just reaped one of the hidden rewards of buying nice kit: a feature you didn't realize was important was included, and you now have a little "free" upgrade! The faster memory speed should slightly speed up regular tasks too, but for Prime95 (and LLR if you search for primes of other forms later) you've seen how that 3600 feature is essential.

You *might* still saturate 3600 speed memory, so you may still benefit from turning boost down a little bit- lower power use and lower temps at "no" cost to P95 speed. I'm considering buying a 5900 or 5950 soon myself, so I am curious if you can quantify your speed improvement from turning XMP on. E.g. what were your old iteration times, and new?

Edit: I am also curious how cinebench and P95 temps are now that memory is feeding the CPU more fully. If you don't mind....
(emphasis added)
I'd wait until the 5800x3D reviews are in; the 96MB L3 might be very interesting. Scheduled release on 4/20-ish
sdbardwick is online now   Reply With Quote
Old 2022-03-22, 03:35   #8
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

2×7×13×29 Posts
Default

I am doubtful that the extra cache will matter for my typical workloads of GMP-ECM and CADO-NFS. I'd rather have extra cores. I do run some LLR, but the FFT sizes fit in any modern CPU cache already (e.g. 256k FFT, exponents around 3-4M).
VBCurtis is offline   Reply With Quote
Old 2022-03-22, 18:04   #9
MarkVanCoutren
 
May 2021

5 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
Cool! You just reaped one of the hidden rewards of buying nice kit: a feature you didn't realize was important was included, and you now have a little "free" upgrade! The faster memory speed should slightly speed up regular tasks too, but for Prime95 (and LLR if you search for primes of other forms later) you've seen how that 3600 feature is essential.

You *might* still saturate 3600 speed memory, so you may still benefit from turning boost down a little bit- lower power use and lower temps at "no" cost to P95 speed. I'm considering buying a 5900 or 5950 soon myself, so I am curious if you can quantify your speed improvement from turning XMP on. E.g. what were your old iteration times, and new?

Edit: I am also curious how cinebench and P95 temps are now that memory is feeding the CPU more fully. If you don't mind....
My pre-XMP iteration times were around 3 ms/iter and now that XMP is enabled, they're usually below 2 (1.8 or 1.9 when I'm not running anything else). Temps went from leveling off at around 71 to around 68 while running prime95 but that could be because of a difference in air temps.

Cinebench is the most interesting part. I ran it yesterday after getting XMP on and discovered the score dropped by around 10,000 points (~25,000 to ~15,000). I spent some time troubleshooting and it might be something related to power delivery. I called MSI support and the guy pretty much said "huh, that's weird" and not much else. I ran the test again with XMP on and the score was 23,000. Nothing about the system changed from yesterday to today so I have no clue what happened. My Cinebench temps were around 66 this run but they were much lower yesterday when I had the issues.
MarkVanCoutren is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Few questions Tumeez Information & Answers 7 2014-09-27 10:40
Many (gmp-)ecm questions... WraithX GMP-ECM 26 2014-04-03 15:54
Some questions... OmbooHankvald PSearch 3 2005-09-17 19:29
5 questions OmbooHankvald Factoring 6 2005-08-28 19:31
Questions OmbooHankvald Prime Sierpinski Project 2 2005-08-01 20:18

All times are UTC. The time now is 12:59.


Thu May 19 12:59:08 UTC 2022 up 35 days, 11 hrs, 0 users, load averages: 1.97, 1.80, 1.74

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.

≠ ± ∓ ÷ × · − √ ‰ ⊗ ⊕ ⊖ ⊘ ⊙ ≤ ≥ ≦ ≧ ≨ ≩ ≺ ≻ ≼ ≽ ⊏ ⊐ ⊑ ⊒ ² ³ °
∠ ∟ ° ≅ ~ ‖ ⟂ ⫛
≡ ≜ ≈ ∝ ∞ ≪ ≫ ⌊⌋ ⌈⌉ ∘ ∏ ∐ ∑ ∧ ∨ ∩ ∪ ⨀ ⊕ ⊗ 𝖕 𝖖 𝖗 ⊲ ⊳
∅ ∖ ∁ ↦ ↣ ∩ ∪ ⊆ ⊂ ⊄ ⊊ ⊇ ⊃ ⊅ ⊋ ⊖ ∈ ∉ ∋ ∌ ℕ ℤ ℚ ℝ ℂ ℵ ℶ ℷ ℸ 𝓟
¬ ∨ ∧ ⊕ → ← ⇒ ⇐ ⇔ ∀ ∃ ∄ ∴ ∵ ⊤ ⊥ ⊢ ⊨ ⫤ ⊣ … ⋯ ⋮ ⋰ ⋱
∫ ∬ ∭ ∮ ∯ ∰ ∇ ∆ δ ∂ ℱ ℒ ℓ
𝛢𝛼 𝛣𝛽 𝛤𝛾 𝛥𝛿 𝛦𝜀𝜖 𝛧𝜁 𝛨𝜂 𝛩𝜃𝜗 𝛪𝜄 𝛫𝜅 𝛬𝜆 𝛭𝜇 𝛮𝜈 𝛯𝜉 𝛰𝜊 𝛱𝜋 𝛲𝜌 𝛴𝜎𝜍 𝛵𝜏 𝛶𝜐 𝛷𝜙𝜑 𝛸𝜒 𝛹𝜓 𝛺𝜔