mersenneforum.org  

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

Reply
 
Thread Tools
Old 2022-01-11, 23:14   #23
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

23×32×41 Posts
Default

Quote:
Originally Posted by Xyzzy View Post
If you put the same $ into a cheap system with a good GPU, which would end up being more efficient?

The problem is memory bandwidth. One channel of DDR4 is good for two cores. Systems with four or eight channels of memory are far more expensive.

When I bought my 4 node i5-6600 cluster in June 2016, I paid $40 per motherboard, $290 per CPU, 120 per 32 GB DDR4 (2 x 16 GB dual rank), shared one $95 power supply, and $75 in power supply cables. It totalled to about CA$2100 after tax (about US$1600 at the time). I got 16 cores and 8 channels of memory. The whole cluster consumed about 270 watts once I disabled boost. On a 4096k FFT, I got 0.466 iter/sec/USD.

You may be able to build an 8 channel 16 core single system for about that in today's prices. An Epyc 7313P costs about US$900. The cheapest memory I could find is $216 fo 8 * 8 GB. An AsRock Rack EPYCD8 ATX motherboard costs $470. A suitable power supply is probably close to $100 (I'd buy at least a Gold, as the power savings are worth it over time). Mind you, that's without the taxes I paid.
Mark Rose is offline   Reply With Quote
Old 2022-01-12, 04:07   #24
axn
 
axn's Avatar
 
Jun 2003

5×29×37 Posts
Default

Quote:
Originally Posted by Mark Rose View Post
The problem is memory bandwidth.
You can remove the dependency on memory bandwidth if you have sufficiently large L3. 64MB L3 of 5900X can hold current wavefront FFT in its entirety.
axn is offline   Reply With Quote
Old 2022-01-12, 09:20   #25
JWNoctis
 
"J. W."
Aug 2021

1000112 Posts
Default

Quote:
Originally Posted by axn View Post
You can remove the dependency on memory bandwidth if you have sufficiently large L3. 64MB L3 of 5900X can hold current wavefront FFT in its entirety.
Is the inter-chiplet data path any faster than main memory, though? Or would that not matter much?

Would make for an interesting comparison versus the upcoming 5800X3D with that huge single-chiplet 96MB L3, though.

As for SFF PC used for number crunching, I could report that a quad-core i7-1165G7 in a NUC maxed out quite a bit slower than an octa-core 5800H in a gaming laptop with the exact same RAM bandwidth, and made quite the ruckus at full power. I'm not sure how long it - fan, VRM, processor silicon, and such - would have lasted at 90+ degC even with cooling at maximum, either.

RAM is probably not nearly the only factor, and I could probably have built something not much larger with a faster CPU and a mini-ITX motherboard for less than that NUC, except for the insane price of any useful graphic card these days.
JWNoctis is offline   Reply With Quote
Old 2022-01-12, 10:09   #26
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

23×32×41 Posts
Default

Quote:
Originally Posted by JWNoctis View Post
Would make for an interesting comparison versus the upcoming 5800X3D with that huge single-chiplet 96MB L3, though.
I'm tempted to buy a 5800X3D. I suspect it will be a monster. It may be the chip to get if the price is reasonable.

As far as I know, with multiple CCX, L3 victims aren't written between CCX, so unless data happens to be in the other L3 cache (from other worker threads), the read must come from memory. So the 64 MB of L3 doesn't really act as a single 64 MB cache.
Mark Rose is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Mini ITX with LGA 2011 (4 memory channels) bgbeuning Hardware 7 2016-06-18 10:32
Mini ITX in server case bgbeuning Hardware 8 2016-05-11 14:13
Mini STX bgbeuning Hardware 1 2016-04-20 13:35
What can I fit in mini-ITX and a 120W PSU fivemack Hardware 2 2016-01-01 17:50
Mini-Projects MrHappy Data 56 2012-09-06 02:48

All times are UTC. The time now is 03:07.


Thu May 26 03:07:07 UTC 2022 up 42 days, 1:08, 0 users, load averages: 1.12, 1.22, 1.18

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