dual rank vs single rank performance and oc capability, also interested in "tri rank" aka mismatched capacity w same ic

[Somerandomtechyboi]

Basically x58 is garbage at clocking rams topping out at ~2200mhz, now im looking for performance in other forms, ive mainly focused on timings cause better timings = lower latency = better performance but i havent really looked into dual rank ram since id rather not have the freq lower than 2200mhz cause thats gonna nuke the benifit alltogether for all i know

 

Now if i dont just sell this thing off and go straight to ryzen ill prob upgrade the rams to 12gb later on so now im faced with the choice of single rank vs dual rank performance and clock capability wise

 

 

For all i know dual rank at lower speed can match or beat single rank at higher speed but idk by how much dual rank performs better than single rank, ics wise im stuck between samsung 4gbit rev b/d/q and micron 4gbit d9qbj, both looking like pretty horrible ics compared to something like samsung 2gbit rev d, cant go hynix 4gbit b/mfr cause again x58 suck at clocking

 

The obvious choice here is dual rank 2gbit rev d due to its good performance characteristics, though i may also be interested in getting 18gb w a 4x3 + 2x3 ram config which may be a "tri rank" config if that even exists

 

I will degradation test both the imc and my soon to come 1x3 micron D9JNx/KPx rams at 1.55-1.6v vtt and 2.46v dram so i may be comfortable with more or less voltages, i expect to run 1.45-1.5v vtt and 2.2-2.46v dram for daily atm

 

 

 

For the questions

How much performance benifit is there from single rank to dual rank? Is it a bandwidth based benifit or latency benifit?

 

How much harder is dual rank to drive compared to single rank? Atm only comfortable with max 1.5v vtt for daily so dont wanna needlessly crank it to 1.6v or something

 

How much freq capability is lost going for dual rank dimms? Cause looking at ddr4 you can run decent bin single rank b die to 5000+ if you tune it right even on a meh board but dual rank looks to have trouble even getting over 4200mhz, there are boards optimized for dual rank but i doubt this first gen ddr3 board has any cause its early gen and add to that x58 imc suck ass, though freq capability also depends on pcb iirc but im going for generic rams so pcbs gonna be trash, Doesnt really matter unless i cant run 2200mhz

 

Regarding "tri rank" does such a thing even exist? Are there any performance penalties with it? And does running mismatched capacity with the same ic nuke ocing capability?

[cmndr]

x58 is a tri-channel set up. For context if you ran RAM on it at 2200MHz, it's be comparable to 3300Mhz on a dual channel set up in terms of bandwidth. Chances are memory is NOT the limit on that old CPU. Modern systems that run DDR4 at 3300MHz or so don't get big bumps from going faster (a few percent at most) and those CPUs have ~40% higher IPC and 10% higher clock speeds and 1-2.7x as many cores to feed. Also new instructions like AVX2 that drastically improve CPU performance (and thus need more memory bandwidth to keep the cores fed).

Also the memory controller is on the CPU, not the chipset. As an FYI, if you're not using a 6 core Xeon CPU, buy one, they're like $10-15 on ebay.


Most of the benefit of having 2 ranks per channel is that each RAM rank refreshes at different times, this means that on average you're able to get slightly better latency. I suspect that with 3 channels a good chunk of this benefit will be diminished since each channel could conceivably refresh at a different time from the others. This is not confirmed.

 

Also... memory performance usually doesn't matter THAT much unless you have really bad RAM. Think DDR3-1333 with awful timings.

 

If you're experimenting for fun, I won't speak out against that but do be aware that you're in the domain of "this REALLY REALLY doesn't matter." 100MHz more on the CPU core will probably matter 2-10x as much as another 100MHz on the RAM.

 

 

Quote

Regarding "tri rank" does such a thing even exist?

Not that I'm aware of. At least not on consumer CPUs.

 

Triple channel RAM has been a thing since... a long time ago. Same goes for quad-channel. I think these days motherboards usually support only dual or quad, though I'd need to confirm. x58 is DEFINITELY triple channel though. x79 moved to quad channel.

If you want a good read check this out - https://www.intel.com/content/www/us/en/support/articles/000005657/boards-and-kits.html

[Somerandomtechyboi]

37 minutes ago, cmndr said:

if you're not using a 6 core Xeon CPU, buy one, they're like $10-15 on ebay.

I have an x5660 but its trash bin (3rd core is complete garbage) so ill (ab)use it for degradation testing and swap for an e5649

 

38 minutes ago, cmndr said:

Also the memory controller is on the CPU, not the chipset

The chipset is linked to ioh while vtt is linked to the imc, ioh only needs 1.3v it seems so its running pretty cool even w/o fan (yes repasted)

 

39 minutes ago, cmndr said:

If you're experimenting for fun, I won't speak out against that but do be aware that you're in the domain of "this REALLY REALLY doesn't matter." 100MHz more on the CPU core will probably matter 2-10x as much as another 100MHz on the RAM.

Im an oc nut so i exctract every bit of performance i can, plus rams usually dont consume much power anyways so no reason not to ram overclock, even upto 2.46v

[porina]

2 hours ago, Somerandomtechyboi said:

How much performance benifit is there from single rank to dual rank? Is it a bandwidth based benifit or latency benifit?

In my testing using Aida64, the difference between 1R and 2R is hardly noticeable. Bandwidth goes up slightly, latency goes up slightly. In a ram demanding application like Prime95 large FFT, I saw +25% performance increase going from 1R to 2R at the same ram speed and timings. Prime95 at big FFT has pretty intense read/write mix and it seems the additional rank really helps out there. Most applications that aren't so dependant on ram performance might see little to no difference. 

 

I didn't test in depth due to hardware on hand, but going beyond 2R/channel didn't seem to offer further benefit.

[Somerandomtechyboi]

45 minutes ago, porina said:

In my testing using Aida64, the difference between 1R and 2R is hardly noticeable. Bandwidth goes up slightly, latency goes up slightly. In a ram demanding application like Prime95 large FFT, I saw +25% performance increase going from 1R to 2R at the same ram speed and timings. Prime95 at big FFT has pretty intense read/write mix and it seems the additional rank really helps out there. Most applications that aren't so dependant on ram performance might see little to no difference. 

 

I didn't test in depth due to hardware on hand, but going beyond 2R/channel didn't seem to offer further benefit.

i guess 2r can be benificial depending on the application, though how do you guage performance increase with prime95? I usually only use it for torture testing

 

as long as 2r dont do under 2000 i think ill chose 2gbit rev d over something like 4gbit rev b/d9qbj

[porina]

50 minutes ago, Somerandomtechyboi said:

how do you guage performance increase with prime95? I usually only use it for torture testing

In the enthusiast community Prime95 might be best known as a torture test but it is doing a real workload and has a built in benchmark function. However you can't just run it and get a single number out to compare. The results will depend on the chosen configuration.

 

The amount of data being worked on depends on the FFT size. To use it to test ram you need to make sure the work is too big to fit in CPU cache and therefore will be impacted by ram. Take FFT size and multiply by 8 to get it to MB. e.g. 1024k (1M) FFT = 8 MB data set. Then multiply again by the number of simultaneous tasks (called workers). Say you have 6 cores, it may default to testing 1, 2, 3, 6 workers. Let it do so, and find the one with the highest iter/s. It can vary due to non-perfect thread scaling. Higher is better. Change your ram settings and try again. For ram testing ignore results where it can fit in CPU cache as it will be much higher performing than when ram limited. Make sure to use the same FFT size, as that will affect performance and you can't compare different FFT sizes. I suggest 1024k as a starting point although you can try higher like 2x and 4x that, especially for Ryzen with bigger caches.

 

To save a bit of time when running the benchmark you can uncheck the test HT box, and reduce the tests to 5s which seems to be the minimum allowed. There will be some run to run variation so run it a few times to check.

 

Of course, this only represents performance running Prime95 and other similar compute workloads.

[Guest]

This guy - You crack me up. 

 

We are on countless threads of the same subject over and over.

 

Prime95 is a stress test. So degredation testing it is!!!

 

Like I've said, again now, again later.

Bench PiMod or Perhaps PiFast so you can get a measurement that's meaningful.

Maybe take some screen shots.

 

Try lower speeds and voltages. 

 

Max voltage, max clock, max stress testing, but nothing has been accomplished in weeks and weeks of new threads and meaningless posts.

 

What exactly is the goal? Bench the same stick of ram to death on prime95????

[Somerandomtechyboi]

23 minutes ago, Guest 5150 said:

Bench PiMod or Perhaps PiFast so you can get a measurement that's meaningful.

Any other memory benchmarks i can try after i run these?

 

24 minutes ago, Guest 5150 said:

Try lower speeds and voltages. 

Haha no

 

24 minutes ago, Guest 5150 said:

What exactly is the goal? Bench the same stick of ram to death on prime95????

Just looking for info on ram configs so i have an idea of what ill buy and what ill buy after that, the 1x3 is already determined D9GTx/JNx/KPx or elpida bdse is what i can find but im more compelled to go micron cause volt scaling

 

Im just asking this so i know what to buy if i do try dailying this 12 year old nuclear reactor, cause id rather not end up with underperforming or garbage ram sticks/config, no i am not gonna lower volts till i see noticable degradation within a short period of time like a week

 

Goal is just get a good/ideal ram config, ideal is preffered but unfortunately sometimes you have to settle for decent ics and not the golden ones like elpida hypers, pretty much just having fun overclocking my stuff to the ground, if they survive then i know what volts i can run and reccomend for x58

 

 

im gonna do the degradation test when the folding comp arrives so i have an excuse to nuke my parents powerbill xD

[freeagent]

6 hours ago, Somerandomtechyboi said:

Basically x58 is garbage at clocking rams topping out at ~2200mhz,

On shitty Xeons maybe. Grab yourself a Gulftown.

[Guest]

10 minutes ago, Somerandomtechyboi said:

Any other memory benchmarks i can try after i run these?

 

Haha no

 

Just looking for info on ram configs so i have an idea of what ill buy and what ill buy after that, the 1x3 is already determined D9GTx/JNx/KPx or elpida bdse is what i can find but im more compelled to go micron cause volt scaling

 

Im just asking this so i know what to buy if i do try dailying this 12 year old nuclear reactor, cause id rather not end up with underperforming or garbage ram sticks/config, no i am not gonna lower volts till i see noticable degradation within a short period of time like a week

 

Goal is just get a good/ideal ram config, ideal is preffered but unfortunately sometimes you have to settle for decent ics and not the golden ones like elpida hypers, pretty much just having fun overclocking my stuff to the ground, if they survive then i know what volts i can run and reccomend for x58

 

 

im gonna do the degradation test when the folding comp arrives so i have an excuse to nuke my parents powerbill xD

You can get much faster systems to better spend your money on, but I think you said you have Ivy Bridge??

 

1366 overclocking, you want a really good board. 

 

Then start binning processors (IMCs).

 

If you where in the states, I'd send you some 1366 hardware.

[freeagent]

IB would be more satisfying to you I think. As long as you have a good CPU 2200 and up should be doable. 2200 is about the limit for my particular turd of a 3770K. I could probably squeeze a bit more, but it wont be stable.

[Somerandomtechyboi]

11 minutes ago, freeagent said:

On shitty Xeons maybe. Grab yourself a Gulftown.

Isnt an x5660 gulftown? What should i get then? Im planning to get an e5649

 

10 minutes ago, Guest 5150 said:

You can get much faster systems to better spend your money on, but I think you said you have Ivy Bridge??

This is more of a fun project, and for all i know a 2600k at 5g will pale in comparison to a 6 core westmere at 5g just because more cores

 

1 minute ago, freeagent said:

IB would be more satisfying to you I think. As long as you have a good CPU 2200 and up should be doable. 2200 is about the limit for my particular turd of a 3770K. I could probably squeeze a bit more, but it wont be stable.

im getting ivy bridge just as a "buisness venture" i guess, basically modding xmp of garbage ddr3 ram to make them cheap alternatives to overpriced 2133/2400 rams, or just outright obliterating them and running higher speeds and tighter timings, considering most of em are at a measly 1.65v, running mine at 2v will prob win by a landslide

 

And ofc ill prob vmod vdimm cause 2.1/2.2 is pathetic, not enough to really squeeze the mostout of my rams, maybe 2.4/2.5v assuming the imc doesnt die, but then again im buying pentiums so its not like i care if i kill one when theyre ~2$ ea

[freeagent]

Oh, so in your opinion every DDR3 IC should be able to run 2133 and up with relatively tight timings? And just an SPD change should correct that?

 

Interesting.

 

We are now moving into DDR5, and DDR4 is considered old now. So you can achieve good latency with DDR3, it can never compete in bandwidth.

[freeagent]

1 hour ago, Somerandomtechyboi said:

Isnt an x5660 gulftown? What should i get then?

Xeons are Westmeres. They don’t have memory dividers like Gulftown does. 970, 980x 990x is what you want for memory.

[Somerandomtechyboi]

39 minutes ago, freeagent said:

Oh, so in your opinion every DDR3 IC should be able to run 2133 and up with relatively tight timings? And just an SPD change should correct that?

No i search for decent ics to first overclock then when i find a decent stable 1.8-2v oc edit the spd to that overclock then sell off

 

It just so happens this rubbish rev f can run 2200 for whatever reason

[freeagent]

I was able to do 2350 on a ud5, 970, and some adatas. I think that was with only 1.7vdimm and not a lot of effort because I was new. Timings were crap I think, but still not terrible.

[Guest]

51 minutes ago, freeagent said:

I was able to do 2350 on a ud5, 970, and some adatas. I think that was with only 1.7vdimm and not a lot of effort because I was new. Timings were crap I think, but still not terrible.

I had some Adata ddr3 also. It was pretty good stuff man. D9 I'm certain. 

 

2000mhz DDR3 as tight as you can get it.

That's where the performance is at. CL6 anyone?

 

DDR4 4000mhz as tight as you can get it. CL12 anyone?

 

DDR5, still in the works. ETA 3 years till perfection, but 8000mhz CL36 would be a good aim.