mersenneforum.org  

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

Reply
 
Thread Tools
Old 2020-04-17, 21:11   #12
PhilF
 
PhilF's Avatar
 
Feb 2005
Colorado

523 Posts
Default

It isn't really different phases.

The utility transformer most US homes are hooked to is a single phase transformer, with a 240V center-tapped secondary. The center tap is grounded and labeled neutral. So half of the outlets in the house are wired from one side of the transformer to neutral, which gives you 120V, and the other half of the outlets are wired to the other side of the transformer to neutral. If you take two outlets that are wired to opposite halves of this split-phase, then use a special cord wired as sdbardwick described, you will indeed get 240V.

But unlike him, I actually have done such a thing, lol.
PhilF is offline   Reply With Quote
Old 2020-04-17, 21:11   #13
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

2×4,643 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
I thought most computer power supplies have a switch (or auto-detect?) to run 240 vs 120V;
It depends on the kit. If it will take ~240, it's more efficient.

Quote:
Originally Posted by VBCurtis View Post
...is it not valid that in this dreamworld I could just wire a regular-looking USA plug (well, six) to the 240V circuit, and run computer power supplies to said plugs without splitting the phase/voltage down to 120+ground?
Don't invite an electrical inspector around to see your cool (or, not so cool) stack! There's a reason different types of plugs and sockets exist. Amongst many things, not having something expecting ~120 being plugged into the outlet by an employee, SWMBO, etc. To do it correctly, you'd use something like a NEMA 2-20 or 2-30 et al.

P.S. For real server power fun, spend some time working with 48V DC systems. But, importantly, remember that the feed is actually -48 volts! It can be an expensive mistake to forget that (been there; observed that!).
chalsall is offline   Reply With Quote
Old 2020-04-17, 21:14   #14
retina
Undefined
 
retina's Avatar
 
"The unspeakable one"
Jun 2006
My evil lair

10110100111002 Posts
Default

Quote:
Originally Posted by PhilF View Post
It isn't really different phases.

The utility transformer most US homes are hooked to is a single phase transformer, with a 240V center-tapped secondary. The center tap is grounded and labeled neutral. So half of the outlets in the house are wired from one side of the transformer to neutral, which gives you 120V, and the other half of the outlets are wired to the other side of the transformer to neutral. If you take two outlets that are wired to opposite halves of this split-phase, then use a special cord wired as sdbardwick described, you will indeed get 240V.

But unlike him, I actually have done such a thing, lol.
Okay, so that is really just a single phase system, with a doubler at the entry point. So how do you run your three-phase A/C?
retina is online now   Reply With Quote
Old 2020-04-17, 21:18   #15
PhilF
 
PhilF's Avatar
 
Feb 2005
Colorado

523 Posts
Default

Quote:
Originally Posted by retina View Post
Okay, so that is really just a single phase system, with a doubler at the entry point. So how do you run your three-phase A/C?
There's no doubling.

But I suppose 3-phase A/C probably would be needed to cool such a monster anyway ...
PhilF is offline   Reply With Quote
Old 2020-04-18, 00:06   #16
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

713310 Posts
Default

This thread has wandered off track. Does any one hae any Ryzen experience to suggest to the OP how his proposed purchase might perform with that big L3 cache?
Prime95 is online now   Reply With Quote
Old 2020-04-18, 06:23   #17
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

211758 Posts
Default

Quote:
Originally Posted by chalsall View Post
Remember that a 240V (more likely 208V) circuit will be two joined phases
Nope. The 240 in most of the world is what 120 is on that side of the pool, i.e. single phase. Joined phases here is 380, etc. Almost all (single phase) computer power supplies we have seen in our life had either a manual switch or auto-selection between 220/240 and 120.

Last fiddled with by LaurV on 2020-04-18 at 06:40
LaurV is offline   Reply With Quote
Old 2020-04-18, 11:53   #18
Xyzzy
 
Xyzzy's Avatar
 
"Mike"
Aug 2002

769910 Posts
Default

Quote:
Originally Posted by Prime95 View Post
This thread has wandered off track. Does any one hae any Ryzen experience to suggest to the OP how his proposed purchase might perform with that big L3 cache?
We have an older and smaller TR system with quad-channel memory that scales poorly with Mprime. Won't more cores just scale even worse?

https://www.mersenneforum.org/showthread.php?t=25024

Xyzzy is offline   Reply With Quote
Old 2020-04-18, 16:59   #19
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

2×5×19×23 Posts
Default

Quote:
Originally Posted by Xyzzy View Post
We have an older and smaller TR system with quad-channel memory that scales poorly with Mprime. Won't more cores just scale even worse?

https://www.mersenneforum.org/showthread.php?t=25024

If the FFT data all fits in CPU cache, it won't really hit main memory and the scaling might be awesome. So, the answer for that will depend on what size exponent is being tested, and just how well cache-accesses work when most or all of the cores are trying to use it. 128MB of cache should hold a first-time-test dataset with quite a bit of room to spare (in fact, two copies if I recall the proper 6*FFTsize = RAM need).
VBCurtis is online now   Reply With Quote
Old 2020-04-18, 22:04   #20
Viliam Furik
 
Jul 2018
Martin, Slovakia

33·7 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
If the FFT data all fits in CPU cache, it won't really hit main memory and the scaling might be awesome. So, the answer for that will depend on what size exponent is being tested, and just how well cache-accesses work when most or all of the cores are trying to use it. 128MB of cache should hold a first-time-test dataset with quite a bit of room to spare (in fact, two copies if I recall the proper 6*FFTsize = RAM need).
Based on my observation, that 3900X with 64 MB of L3 cache (if the L2 cache counts too, then it's 64+6 MB), can hold two tests, having FFT size at most 2880K, and thus occupying 32MB, I think the correct formula is 12 * FFT size = memory needed. (12 * 2 880 000 = 34 560 000 B).
Viliam Furik is offline   Reply With Quote
Old 2020-04-19, 02:42   #21
axn
 
axn's Avatar
 
Jun 2003

22×11×107 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
(in fact, two copies if I recall the proper 6*FFTsize = RAM need).
Quote:
Originally Posted by Viliam Furik View Post
I think the correct formula is 12 * FFT size = memory needed.
Try 8*FFT size (+ overhead).
axn is offline   Reply With Quote
Old 2020-04-19, 08:24   #22
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

1100001012 Posts
Default

Quote:
Originally Posted by axn View Post
Try 8*FFT size (+ overhead).
I found this works well without overhead... at least on lower core count systems (up to 8 or so?). Over on PrimeGrid some users of higher core count Zen 2 models (32+) reported lower than expected performance with multiple LLR instances. Switching from Windows to Linux seemed to help, but I don't think explains the whole story. I find it fascinating to explore but without such a system on hand it isn't something I expect to look at. My hunch is the "+ overhead" starts to become significant with more parallel tasks running on higher core count CPUs.

https://linustechtips.com/main/topic...umber-finding/

Even on 6 core model, the scaling running 1 worker can be seen as a glass half empty/half full scenario. At the above link I visualised the Prime95 benchmark result at 1, 2, 6 workers for a 8086k and 3600. I can't remember the formula to normalise for FFT size, but it had a lot of logs in it. The 1 worker result on 3600 offers best performance at bigger FFTs but I think the relative throughput is lower than expected, which I attribute to the CCX fragmentation. Rumours of Zen 3 having 8 cores as the smallest logical unit compared to 4 cores now make them very interesting.

I don't have compelling evidence if the L2 should be added to L3 in the case of Zen 2. The caches are exclusive but the L2 is relatively small so might not be so significant. Adding L2+L3 does seem appropriate on Skylake-X for example. Performance is better than if you look at L3 alone, but in that case the L2 is closer in total size to the L3 so more significant.
mackerel is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Has anyone tried linear algebra on a Threadripper yet? fivemack Hardware 3 2017-10-03 03:11

All times are UTC. The time now is 05:25.

Wed Oct 21 05:25:47 UTC 2020 up 41 days, 2:36, 0 users, load averages: 1.25, 1.35, 1.43

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.