mersenneforum.org  

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

Reply
 
Thread Tools
Old 2019-01-10, 10:35   #23
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

2×32×5×7 Posts
Default

Quote:
Originally Posted by wildrabbitt View Post
I thought RAM speed only really mattered for trial factoring?

Is it important for LL tests too?
Yes but there's more to it than just speed, AFAIK you should also consider the timings. tl;dr 3000 speed with CL15 or 3200 speed with CL16 is very common for DDR4 and you won't go far wrong as those settings are fairly well supported across motherboards and somewhat at the sweet spot of price to performance. Rule of thumb you're looking for higher speeds without sacrificing too much CL, lower CL is better.
M344587487 is offline   Reply With Quote
Old 2019-01-10, 11:10   #24
axn
 
axn's Avatar
 
Jun 2003

7·11·61 Posts
Default

Quote:
Originally Posted by wildrabbitt View Post
I thought RAM speed only really mattered for trial factoring?

Is it important for LL tests too?
There is no "too". It is only important for LL. TF is not at all sensitive to RAM speed.

EDIT:- Since P-1 & ECM also depend on the same FFT-based calculation that LL does, they're also sensitive to memory speed. Ironically, TF is the only thing that is NOT sensitive to memory speed :)

Last fiddled with by axn on 2019-01-10 at 11:15
axn is offline   Reply With Quote
Old 2019-01-10, 17:58   #25
PhilF
 
PhilF's Avatar
 
Feb 2005
Colorado

22×127 Posts
Default

Quote:
Originally Posted by wildrabbitt View Post
I thought RAM speed only really mattered for trial factoring?

Is it important for LL tests too?
It is very important for any calculations that use a FFT size that is larger than your processor's internal cache size.

Since the size of the FFT is usually larger than the processor's cache, parts of it are constantly being swapped in and out of memory as needed during the calculation. So even though the amount of memory being accessed is relatively small, the raw speed and bandwidth available for that access makes a huge difference in iteration times.
PhilF is offline   Reply With Quote
Old 2019-01-10, 19:17   #26
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

6018 Posts
Default

On ram bandwidth vs latency, in my past testing there was a very strong impact from bandwidth, and a hardly detectable one from latency. Whatever is in place to mask latency is working well. so if buying ram for this application do not sacrifice bandwidth for lower latency, unless you know you are past bandwidth limiting.
mackerel is offline   Reply With Quote
Old 2019-01-11, 15:30   #27
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

2×32×5×7 Posts
Default

Quote:
Originally Posted by mackerel View Post
On ram bandwidth vs latency, in my past testing there was a very strong impact from bandwidth, and a hardly detectable one from latency. Whatever is in place to mask latency is working well. so if buying ram for this application do not sacrifice bandwidth for lower latency, unless you know you are past bandwidth limiting.
That's interesting, I'll have to try scuppering the latency sometime for a little extra speed.
M344587487 is offline   Reply With Quote
Old 2019-02-22, 23:45   #28
mognuts
 
mognuts's Avatar
 
Sep 2008
Bromley, England

1010102 Posts
Default

I've been following this thread with interest. My computer is due for an upgrade, as it's quite old in computer terms:

This is the system I have now:

Intel i5-2500K CPU overclocked to 4500MHz
Asus P8P67 PRO Intel P67 Motherboard
Corsair Vengeance (4x8GB) DDR3 RAM (1866MHz 9-10-9-27)

I am considering the following:

Intel I5-8600K CPU [£240]
Asus Prime Z370-A II Intel Z370 Motherboard [£160]
Team Group Xtreem Edition (2x8GB) DDR4 RAM (4000MHz 18-19-19-39) [£200]

The RAM is exactly the same price as 3600Mhz 16-16-16-36 RAM by the same manufacturer/vendor, so I'm assuming that if 4000MHz is too high for the CPU/Motherboard combo, I can always underclock it to 3600Mhz or whatever it's stable at, without any performance or financial penalty. The RAM uses Samsung B-dies, and I believe it’s single rank.

I shall be using the system for GIMPS (LL/DC/PRP, and occasional P-1) about 70% of the time, and general computing stuff for the remainder.

I don’t intend to overclock the CPU for GIMPS, but I may fiddle with the settings for video rendering, games etc.

Can anybody suggest a better system for the total outlay (£600), or have any opinions on individual components etc. Any advice and comments gratefully received.


John
mognuts is offline   Reply With Quote
Old 2019-02-23, 01:08   #29
PhilF
 
PhilF's Avatar
 
Feb 2005
Colorado

22·127 Posts
Default

I just purchased the following, by judicious shopping:

Intel i7-8700 - $275.00
(2) V-Color 8GB DDR4-3200 RAM - $108.00 total
GIGABYTE Z390 Gaming SLI Motherboard - $160.00

That's a total of $543.00, including shipping. Notice the CPU is not the "K" version, so overclocking isn't possible. However, what you get in return is tremendous performance with only 65 Watts TDP (the 8700K is 95 Watts).
PhilF is offline   Reply With Quote
Old 2019-02-23, 01:16   #30
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

54538 Posts
Default

Quote:
Originally Posted by mognuts View Post
I've been following this thread with interest. My computer is due for an upgrade, as it's quite old in computer terms:

This is the system I have now:

Intel i5-2500K CPU overclocked to 4500MHz
Asus P8P67 PRO Intel P67 Motherboard
Corsair Vengeance (4x8GB) DDR3 RAM (1866MHz 9-10-9-27)

I am considering the following:

Intel I5-8600K CPU [£240]
Asus Prime Z370-A II Intel Z370 Motherboard [£160]
Team Group Xtreem Edition (2x8GB) DDR4 RAM (4000MHz 18-19-19-39) [£200]

The RAM is exactly the same price as 3600Mhz 16-16-16-36 RAM by the same manufacturer/vendor, so I'm assuming that if 4000MHz is too high for the CPU/Motherboard combo, I can always underclock it to 3600Mhz or whatever it's stable at, without any performance or financial penalty. The RAM uses Samsung B-dies, and I believe it’s single rank.

I shall be using the system for GIMPS (LL/DC/PRP, and occasional P-1) about 70% of the time, and general computing stuff for the remainder.

I don’t intend to overclock the CPU for GIMPS, but I may fiddle with the settings for video rendering, games etc.

Can anybody suggest a better system for the total outlay (£600), or have any opinions on individual components etc. Any advice and comments gratefully received.


John
Looks well-balanced to me. You will probably get a 10% boost in the future when you populate the other RAM slots.

Looking at PC Part Picker, you can get a 9600K a few quid cheaper.

The only other thing I'd add is an 80Plus Gold rated power supply if you don't already have one. Its efficiency will pay for itself in a year or two with UK electricity prices, especially if you also use your GPU for GIMPS.
Mark Rose is offline   Reply With Quote
Old 2019-02-23, 10:56   #31
Clayson
 
Feb 2019

116 Posts
Default

10% boost in what regard btw? The overall performance of the machine?
Clayson is offline   Reply With Quote
Old 2019-02-23, 17:36   #32
PhilF
 
PhilF's Avatar
 
Feb 2005
Colorado

22×127 Posts
Default

Quote:
Originally Posted by Clayson View Post
10% boost in what regard btw? The overall performance of the machine?
He is referring to memory throughput, although I think 10% is optimistic. Regardless, whether this 10% improvement in memory throughput is meaningful depends on what the machine is being used for. Most would never notice, but if the machine is running Prime95 the improvement can be noticeable.

The reason for the improvement in throughput is because of "interleaving". The memory is arranged such that sequential memory locations are physically interleaved across all the populated memory slots. This allows each DIMM to perform its required (and relatively slow) refresh cycle at the same time a different DIMM is being being accessed. Since most memory access is sequential in nature, populating four (or more) DIMM slots can improve overall throughput compared to populating just two, because the CPU is wasting less time waiting for memory refresh cycles to complete.
PhilF is offline   Reply With Quote
Old 2019-02-23, 20:58   #33
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

54538 Posts
Default

Quote:
Originally Posted by PhilF View Post
He is referring to memory throughput, although I think 10% is optimistic. Regardless, whether this 10% improvement in memory throughput is meaningful depends on what the machine is being used for. Most would never notice, but if the machine is running Prime95 the improvement can be noticeable.

The reason for the improvement in throughput is because of "interleaving". The memory is arranged such that sequential memory locations are physically interleaved across all the populated memory slots. This allows each DIMM to perform its required (and relatively slow) refresh cycle at the same time a different DIMM is being being accessed. Since most memory access is sequential in nature, populating four (or more) DIMM slots can improve overall throughput compared to populating just two, because the CPU is wasting less time waiting for memory refresh cycles to complete.
Exactly. The 10% is an approximation. I've seen higher and lower benefit from adding additional memory ranks.
Mark Rose is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Hardware robert44444uk Prime Gap Searches 45 2018-03-12 23:59
NAS hardware VictordeHolland Hardware 5 2015-03-05 23:37
Hardware error Citrix Prime Sierpinski Project 12 2006-06-07 09:40
Hardware Of the Week #1 moo Hardware 4 2005-10-19 15:58
specialized hardware tjdziuba Hardware 1 2004-06-21 17:42

All times are UTC. The time now is 18:35.

Thu Sep 24 18:35:38 UTC 2020 up 14 days, 15:46, 0 users, load averages: 1.48, 1.97, 2.05

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, 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.