mersenneforum.org  

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

Reply
 
Thread Tools
Old 2016-03-20, 22:41   #1
Fred
 
Fred's Avatar
 
"Ron"
Jan 2016
Fitchburg, MA

11000012 Posts
Default i3 w/DDR4?

Anyone running an i3-6100, 6300 or 6350 with DDR4-2133?

All of the following assumes dedicated build running solely to crunch primes.

I'm really curious about how an i3 with DDR4-2133 would benchmark. In particular, I'm thinking about how with my i5-6500 the 3rd and 4th cores get so gimped by memory bandwidth, even with overclocked DDR4-2133.

Even though the i3 only has 2 cores, would they run with virtually no memory bandwidth issues? Given the price difference between an i3-6100 and an i5-6500 (nearly double the price), might an i3 (2 cores, 3.7Ghz, potentially little memory bandwidth issues) be at all comparable to an i5 (4 cores, 3.2Ghz, with memory bandwidth issues?

I suspect the i3 system would certainly be slower, but by how much? If only 10%-20% slower, it seems the cost difference (at least for up front costs) might make the i3 a contender. Of course, power use would be a whole different factor. For the time being, I have a sponsor where I can run my builds, so power isn't a major concern.

Last fiddled with by Fred on 2016-03-20 at 22:42
Fred is offline   Reply With Quote
Old 2016-03-21, 03:48   #2
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

5·11·53 Posts
Default

I'm running an i3-4170 at 3.7 GHz in a machine with DDR3-1333, and the second core is choked by about 60% (I see only a 40% improvement with two cores, not 100%). Simple math says DDR3-1866 would be the minimum desired, and I suspect 2133 would be just about right.

A motherboard and power supply for an i3 is the same as for an i5. If you're looking for best bang for the buck, you're better off with the i5 and overclocked RAM.

I am also curious how an i5-6400 (3.1 GHz four core turbo) would perform vs the i5-6500 (3.3 GHz four core turbo). It's $15 or so cheaper.

Last fiddled with by Mark Rose on 2016-03-21 at 03:50
Mark Rose is online now   Reply With Quote
Old 2016-03-21, 08:01   #3
henryzz
Just call me Henry
 
henryzz's Avatar
 
"David"
Sep 2007
Cambridge (GMT/BST)

574810 Posts
Default

Wouldn't an i3 be lower power though? Looks like you can get 3.7 GHz at 51W and 3.3 GHz at 35W.
This is compared with 65W for the i5s.
I suppose that the question then is whether the hyperthreading of an i3 is worth anything. Would a Pentium be better?
henryzz is offline   Reply With Quote
Old 2016-03-21, 10:50   #4
axn
 
axn's Avatar
 
Jun 2003

17·281 Posts
Default

Quote:
Originally Posted by henryzz View Post
Would a Pentium be better?
Pentium/Celerons lack AVX, so no good. mackerel posted an analysis on PG on the same topic. Perhaps he can chime in.
axn is offline   Reply With Quote
Old 2016-03-21, 19:12   #5
henryzz
Just call me Henry
 
henryzz's Avatar
 
"David"
Sep 2007
Cambridge (GMT/BST)

22·3·479 Posts
Default

Quote:
Originally Posted by axn View Post
Pentium/Celerons lack AVX, so no good. mackerel posted an analysis on PG on the same topic. Perhaps he can chime in.
Bother I read chime as chip first time through. Was about to make a joke about just needing batter.
henryzz is offline   Reply With Quote
Old 2016-03-23, 07:37   #6
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

1100010012 Posts
Default

Late to the thread, but I went through a hypothetical build exercise myself recently. I think the i3-6100 is the performance and value sweet spot assuming you run a separate task per real core.

I didn't go through every CPU combination possible, but roughly speaking i3-6100 + dual channel 2133 ram runs about 90% of its unlimited ram performance. For comparison, i5-6600k at 3.5 GHz, with dual channel 3000 ram would be about 75% of its potential. Since it is in the ram bandwidth limiting zone, it may make more sense to look at lower end i5 models actually.

For the build I'm including a budget SSD, case and PSU. Just the cheapest minimum you can get away with in each case. In the case of the i3 I'm assuming the use of the stock intel cooler, but for the i5 I assume you would want something better. A rough bang per buck metric would put the i3 far ahead. Performance per watt is harder as it would be diluted by the system consumption which I can't predict, but based only on CPU TDP, the i3 above would again be ahead by a fair amount. The beauty of the i3 system is you just look for the cheapest kit around it to get it going, whereas if you go higher end you start paying more attention to other details and the price starts going up more too.

Note the modelling I'm using is manually calculated from past observations. It may be a bit pessimistic when it comes to the performance drop as ram bandwidth starvation kicks in. Still, if I were to build a processing farm today I would base it around the Skylake i3 models. For other projects there might be a small advantage from the bigger L3 cache of higher models.

If supported, running multiple cores on a single task helps a lot. I'm running other prime projects where this isn't possible though.

If more depth is needed on any of the above I can try to expand on it.
mackerel is offline   Reply With Quote
Old 2016-03-23, 15:51   #7
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

157668 Posts
Default

If you buy an i3, please post your throughput results!

I can tell you that in the throughput benchmarks of an i5, running 2 cores gets about 74% of the 4 core throughput (123.49 iter/s vs. 166.64 iter/s).

If iter/s scales with CPU frequency on the i3, you might get up to 123.49 * 3.7 / 3.3 iter/s. Let's say you get 135. That is just 19% less than the i5.

Whether that would make sense in a farm is debatable. Yes, the CPU might cost 50% less, but by the time you buy a mobo, RAM, PSU, the cost savings for a total system will be much less than 50%.
Prime95 is online now   Reply With Quote
Old 2016-03-23, 18:00   #8
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

3·131 Posts
Default

Is it possible to run high speed ram in non-Z170 chipset mobos? I think I saw that somewhere which would help the cost of an i5 system somewhat, but I haven't considered that scenario.

My consideration was that with an i3, you buy the cheapest mobo, ram, everything available. You don't need anything other than the bare minimum spec to get it to run.

With an i5, you need faster ram so as not to totally cripple it, so I was factoring in more expensive ram, mobo, cooling... it isn't just the CPU cost difference. The system cost will go up, and in my estimation performance will not go up faster than the cost of uprating everything.
mackerel is offline   Reply With Quote
Old 2016-03-23, 18:50   #9
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

6118 Posts
Default

Ok, done some digging. It looks like some low cost Asrock models can do some hacked ram overclock to 2800 if you use certain modules. Combine it with i5-6400 and the lowest cost 2800 ram, it worked out about parity to i3-6100 systems by my measure and build assumptions for bang for buck.

So with the uncertainty as to what that offers, the i3 route seems lower risk.
mackerel is offline   Reply With Quote
Old 2016-03-23, 18:57   #10
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

2·3·1,193 Posts
Default

Read the "George's dream build" thread. Lots of data there on i5 builds and bang-for-the-buck. We need i3 throughput numbers to do a full comparison.

P.S. Love the ASrock hacked memory overclock.
Prime95 is online now   Reply With Quote
Old 2016-03-23, 23:44   #11
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

3·131 Posts
Default

I did some testing by clocking and disabling cores to simulate the i3-6100 and i5-6400 at base clock. I presume even as a non-k some mobos can boost 4 cores above base within the turbo limit as I've seen on Haswell. Of course the i7 will have more cache than the others, but past experience showed it made no difference beyond measurement variations in a past comparison at these bigger FFT sizes.

CPU is i7-6700k fixed at 1.2v for all cases. Clock and core count varied. HT off.
Ram is 2x4GB DDR4 run at 2800 16-18-18-36 1.35v and 2133 15-15-15-36 1.2v according to the test. Ram modules are single rank.

Power at the wall was also measured. System uses "82+" PSU (it is a known brand, but I can't see the label in installed position), MSI Gaming Pro (Z170). Coolermaster Hyper 212 Evo for cooling and one vent fan. There's a SSD of some sort in there, as well as a gaming keyboard and random mouse. Onboard graphics used.

Simulated i5-6400 at 2.7 GHz 4 cores

ram 2800
power 94w
Performance per watt: 1.73

Timings for 4096K FFT length (4 cpus, 1 worker): 6.11 ms. Throughput: 163.76 iter/sec.
Timings for 4096K FFT length (4 cpus, 2 workers): 12.46, 12.37 ms. Throughput: 161.07 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 24.62, 24.54, 24.62, 24.71 ms. Throughput: 162.47 iter/sec.


ram 2133
power 91w
Performance per watt: 1.72

Timings for 4096K FFT length (4 cpus, 1 worker): 8.21 ms. Throughput: 121.83 iter/sec.
Timings for 4096K FFT length (4 cpus, 2 workers): 12.94, 12.96 ms. Throughput: 154.43 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 25.45, 25.37, 25.70, 25.64 ms. Throughput: 156.63 iter/sec.


Simulated i3-6100 at 3.7 GHz 2 cores

ram 2800
power 79w
Performance per watt: 1.66

Timings for 4096K FFT length (2 cpus, 1 worker): 7.77 ms. Throughput: 128.77 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 15.27, 15.27 ms. Throughput: 130.98 iter/sec.


ram 2133
power 77w
Performance per watt: 1.58

Timings for 4096K FFT length (2 cpus, 1 worker): 7.93 ms. Throughput: 126.05 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 16.53, 16.36 ms. Throughput: 121.63 iter/sec.


Actual i7-6700k at 4.0 GHz 4 cores
Ram 3000 4x4GB dual rank per channel
Power 126W
Performance per watt: 1.80

Timings for 4096K FFT length (4 cpus, 1 worker): 4.42 ms. Throughput: 226.32 iter/sec.
Timings for 4096K FFT length (4 cpus, 2 workers): 8.89, 8.91 ms. Throughput: 224.73 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 17.48, 17.46, 17.77, 17.72 ms. Throughput: 227.20 iter/sec.

By these benchmark results, the simulated i5-6400 system with "fast" ram is about 34% faster than the simulated i3-6100 with standard DDR4. Obviously I can't put a bang for buck metric easily here since pricing would depend on what you can get locally. So it still comes down to, can you build the i5 system for less than 34% more than the i3 if you consider up front costs only? Those Asrock mobos do certainly help, but I would like to understand more exactly what ram settings are used in the OC mode. Can anyone read out the speed and timing used in that mode, compared to the ram rated speeds?

Considering performance per watt (iter/sec/watt), the i5 system does have just over 9% advantage over i3, but I would caution I've in no way optimised this system for power which might swing things a bit. My fully loaded i7 configuration (how I currently run it) works out even better. Note I don't push CPU clock harder, which is quite easy, as this system is still ram bandwidth restricted.
mackerel is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Does ECC DDR4 work in Haswell-E motherboards fivemack Hardware 2 2015-12-22 04:51
Quad Channel DDR4 vs Dual Channel DDR4 ATH Hardware 17 2015-10-27 21:06
DDR4, Is This Old News? M29 Hardware 6 2014-01-25 22:06

All times are UTC. The time now is 21:58.

Thu Nov 26 21:58:35 UTC 2020 up 77 days, 19:09, 4 users, load averages: 1.79, 1.52, 1.48

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.