Faster RAM, but slower benchmarks?

[Carfax83]

I recently bought this RAM from Newegg. It's the fastest DDR4 3200 RAM on the market to my knowledge, with extra tight timings.

This is the RAM I had before. As you can see, the timings for my old kit are much higher compared to the new one.

However, I benchmarked the new kit using Aida64 Extreme, and I was surprised to see that my scores had actually lowered, with latency being the sole exception

This was the benchmark with my old kit:

And this is the benchmark with my new kit:

What the hell? My read, write and copy are lower with the new set despite the much lower timings. Only latency has improved significantly..

Everything is the same, including the core and uncore frequency which I have set to 3500 max.

One thing I knew beforehand but didn't think it would matter, is that the new kit is actually a DUAL CHANNEL kit, despite having 8GBx4 DIMMs. My original kit was quad channel however.

My theory is that since the dual channel kit has unmatched sets, it's somehow messing with the performance of the kit in quad channel mode. I mean, it's not a huge difference mind you, but it's significant.

Any ideas? I'm going to call G.Skill in the morning and ask them either way though, but this is bugging me.

 

[BonzaiDuck]

Well, all I can say is that your latency in nano-seconds for the new RAM is better than it was for the old RAM.

As I remember, there is a secondary latency which defines the overall delay in clock-cycles for a memory operation and including the sum of clock-cycles for the four primary latencies. It is possible to shorten this secondary latency and improve performance somewhat. I just forgot what it's called . . . . But it might vary with make, model and other features between old RAM and new RAM.

 

[LTC8K6]

North Bridge Clock?

 

[Carfax83]

Well, all I can say is that your latency in nano-seconds for the new RAM is better than it was for the old RAM.

As I remember, there is a secondary latency which defines the overall delay in clock-cycles for a memory operation and including the sum of clock-cycles for the four primary latencies. It is possible to shorten this secondary latency and improve performance somewhat. I just forgot what it's called . . . . But it might vary with make, model and other features between old RAM and new RAM.

You're talking about the row refresh cycle time. I already checked that, and on the older kit it was slightly lower than the new kit. Anyway, I put the same timing in for both kits, and whilst the benchmark scores did improve, it's not the main reason why the read, write and copy are lower.

There's another sub timing setting(s) that I'm missing. Of course, the easiest most pain in the ass thing to do would be to simply reinstall my old kit and then use it as a cross reference for the new one, but that would require uninstalling my video card, and then uninstall my heatsink fans and I'm too lazy to be honest

Anyway, I spoke to G.Skill tech support, and he said that the secondary sub timings are likely different. Since I never really mess with the secondary sub timings, I never even thought to write them down to compare them..

North Bridge Clock?

No, even though the North Bridge Clock aka uncore is showing different speeds in both pics, they are actually the same. That's just an anomaly of Aida64, much like how it's not reporting my CPU frequency properly, which is 4.4ghz and not 3.5ghz.

 

[razel]

Same bandwidth but a difference of 2 in timings. You have decreased your latency which is awesome, but you certainly have wasted your money. It can be better spent on other things... which seems to be a constant truth with RAM.

 

[Carfax83]

Same bandwidth but a difference of 2 in timings. You have decreased your latency which is awesome, but you certainly have wasted your money. It can be better spent on other things... which seems to be a constant truth with RAM.

G.Skill tech support gave me some settings to try out. I haven't finished tweaking them, but here's an example of their impact:

Even lower latency, and everything else except the copy speed is up to par with my original kit. These latency scores are insane for quad channel!

Anyway, you have a misconception as to the value of faster RAM I see, which a lot of people do.

Maybe at one time that was true, but not anymore. As big and as complicated as games are now, RAM speed definitely impacts performance in a significant manner depending on your setup.

Rather than write a lengthy post explaining why, I'll just direct you over to this twenty page thread over at Overclock.net which goes into the nitty gritty in detail.

Quad channel setups like mine benefit less from fast RAM as the CPU has a much larger cache compared to mainstream parts. But even so, it still benefits. The more CPU bound a game is, the greater role RAM frequency plays in performance:

DDR4 2666 is likely the sweet spot for Haswell-E in terms of price and performance. Although I personally can still notice a difference in smoothness when my RAM is running DDR4 2666 and DDR4 3200. At any rate, the main reason why I got this newer kit, is because I like to mess with timings, and this kit gives me the flexibility to do so.

Also, at DDR4 3200 CL14 speed, DDR4 has officially destroyed DDR3 in every conceivable manner, including latency. Even one of the fastest DDR3 kits on the market which runs at 3200, has higher timings than the one I just bought.

I find it amusing, because some people predicated it would take DDR4 at 4000+ speeds to surpass DDR3 when it came to latency

 

[Dasa2]

Also, at DDR4 3200 CL14 speed, DDR4 has officially destroyed DDR3 in every conceivable manner, including latency. Even one of the fastest DDR3 kits on the market which runs at 3200, has higher timings than the one I just bought.

I find it amusing, because some people predicated it would take DDR4 at 4000+ speeds to surpass DDR3 when it came to latency

very nice speeds its good to see that ram working well with x99
sorry but your not there yet different versions of aida report different latency
samsung ddr3 2133 9-10-10 40.2ns

as for why your not seeing a increase in read\write speeds it may be a limitation of the memory controller
going from 2133 to 2400 doesn't increase read\write speed with my 3770k

what settings did gskill have you change to improve the results?

 

[BonzaiDuck]

You're talking about the row refresh cycle time. I already checked that, and on the older kit it was slightly lower than the new kit. Anyway, I put the same timing in for both kits, and whilst the benchmark scores did improve, it's not the main reason why the read, write and copy are lower.

There's another sub timing setting(s) that I'm missing. Of course, the easiest most pain in the ass thing to do would be to simply reinstall my old kit and then use it as a cross reference for the new one, but that would require uninstalling my video card, and then uninstall my heatsink fans and I'm too lazy to be honest

Anyway, I spoke to G.Skill tech support, and he said that the secondary sub timings are likely different. Since I never really mess with the secondary sub timings, I never even thought to write them down to compare them..

I can't feel confident enough for any authority about it, but that timing -- the row refresh cycle if that's what it is -- can be tweaked below the spec setting. Which -- it seems -- you may have done, to report improved benchies.

As I remember for DDR3, you could set the limit on that timing so that it was about two to four clock-cycles greater than TRC : Row Cycle Time. TRC = TRAS + TRP. KEEP IN MIND that my memory is vague about this. It's been a few years since I bothered with it.

But you can tweak it and run HCI Memtest-64 against it until you're sure that it's reliable and stable.

I wouldn't know exactly what sort of performance gains you could expect. When I played with it -- and it's been a few years as I said -- I saw performance gains. But let's be real: the benchmarks tell you one thing; the subjective "feel" of it tells you another.

Note also -- I looked at your memory specs and it shows command-rate = 1 -- which much mean for the XMP setting. If it's set to 2, you should be able run it at 1 given the specs.

 

[razel]

Anyway, you have a misconception as to the value of faster RAM I see, which a lot of people do.

Rather than write a lengthy post explaining why, I'll just direct you over to this twenty page thread over at Overclock.net which goes into the nitty gritty in detail.

We're all still waiting for your not lengthy post... let us know when you're done 'waxing ecstatic' about your purchase. Everyone loves new toys and showing others how cool it is. I really do, but I don't buy things to make myself feel better. Have fun.

 

[Carfax83]

We're all still waiting for your not lengthy post... let us know when your done masturbating about your purchase. Everyone loves new toys and showing others how cool it is. I really do, but I don't buy things to make myself feel better. Have fun.

I'm sure you have better things to do than to troll my thread, or fret about what I spend my money on..

And keep waiting, because I'm not going to waste my time trying to educate someone who can't even take the time to research the subject themselves.

Your post has been reported.. Now run along

 

[Carfax83]

very nice speeds its good to see that ram working well with x99
sorry but your not there yet different versions of aida report different latency
samsung ddr3 2133 9-10-10 40.2ns

That's dual channel though. I have quad channel, which has inherently higher latency due to accessing four memory channels rather than two.

For a quad channel interface, my latency times are extremely low. Back when I had my 4930K rig with quad channel DDR3 2133 CL9, I don't ever recall getting anywhere near a latency of 51..

as for why your not seeing a increase in read\write speeds it may be a limitation of the memory controller
going from 2133 to 2400 doesn't increase read\write speed with my 3770k

No it's not the memory controller. I'm certain of that. It has to do with the secondary and tertiary sub timings, which are looser with the new kit than the old kit. Apparently G.Skill made some sacrifices to achieve the very low primary sub timings.

Although I don't see why they had to do that to be honest. I just ran some stability tests, and even with these new settings, everything is still perfectly stable.

what settings did gskill have you change to improve the results?

They gave me a whole list of stuff. This is what they gave me. So far I've only really switched three settings, which I highlighted in bold. I'll do the rest over time, as I don't want to make my rig unstable.

But to give you an idea of how these new settings compared with the default. The tRFC on default was like 537 if I recall. The new setting is 270.

Why G.Skill had such a high tRFC is beyond me, as the RAM is clearly capable of going much lower..

tWR 15
tRFC 270
tRRD 4
tRRD-L 6
tWTR 3
tWTR-L 0
tRTP 10
tFAW 26
tCKE 6
tCWL 9
tCCD 4
tCCD-L 6
tCCD-WR-L 6
tRWSR 6
tRWDD 4
tRWDR 4
tWRDD 2
tWRDR 1
tWWDD 3
tWWDR 3
tRRDD 1
tRRDR 1
tREFI 30000

 

[Carfax83]

I can't feel confident enough for any authority about it, but that timing -- the row refresh cycle if that's what it is -- can be tweaked below the spec setting. Which -- it seems -- you may have done, to report improved benchies.

Yeah, that was one of the first settings I changed, even before the G.Skill rep gave me those numbers. Although I didn't think to set the tRFC anywhere near as low as 270..

I wouldn't know exactly what sort of performance gains you could expect. When I played with it -- and it's been a few years as I said -- I saw performance gains. But let's be real: the benchmarks tell you one thing; the subjective "feel" of it tells you another.

RAM performance has been so underrated over the years. It can make a big difference depending on your setup. For straight quad cores and dual cores, RAM frequency can result in very large gains, as you can see in this video:

Impact of RAM frequency on various CPUs

As for timings, they can also impact significantly as well, but it's very platform dependent:

Impact of timings on performance in Mafia 2

HT enabled quad cores and above though still benefit, but less, due to larger caches and HT. One of the benefits of HT is that it helps to hide latency, as if one thread stalls, then another can simply take it's place.

Note also -- I looked at your memory specs and it shows command-rate = 1 -- which much mean for the XMP setting. If it's set to 2, you should be able run it at 1 given the specs.

Yep, done that already. One of the first things I do is set my CR to 1..

 

[BonzaiDuck]

. . . .


. . .
Yep, done that already. One of the first things I do is set my CR to 1..

There was a time, maybe just a few years ago, that RAM didn't ship with CR=1 even as an XMP spec. I made sure to test it thoroughly before tweaking the IMC/VCCIO/VTT-whatever voltage or the RAM voltage itself.

With G.SKILL, I found you needn't do either, or if you did, the increase in VCCIO was very small -- from 1.000 to about 1.070V at most.

I used to be very keen on RAM overclocking, and it just adds a lot of extra trouble before you sort it out. For instance, I had some G.SKILL Ripjaws "GBRL" DDR3-1600 sticks, and with the help of another member, ran them at 1866 with looser timings. But why bother, when you can simply buy the RAM spec'd at the speed you want? For hardly chump-change at today's prices?

 

[Dasa2]

That's dual channel though. I have quad channel, which has inherently higher latency due to accessing four memory channels rather than two.

so it does
there is some interesting info here although they use fairly lose timings
http://www.corsair.com/en/blog/2015/september/ddr3_vs_ddr4_generational

 

[Carfax83]

I used to be very keen on RAM overclocking, and it just adds a lot of extra trouble before you sort it out.

Tell me about it. I've corrupted Windows installations because of my penchant for RAM overclocking and messing with sub timings

RAM overclocking is more dangerous than any other type of overclocking in my experience, because the chance for data corruption is much higher. Which I suppose is due to the fact that RAM stores a lot of data.

For instance, I had some G.SKILL Ripjaws "GBRL" DDR3-1600 sticks, and with the help of another member, ran them at 1866 with looser timings. But why bother, when you can simply buy the RAM spec'd at the speed you want? For hardly chump-change at today's prices?

That's exactly why I ended up buying this new kit, as the RAM is far superior to the earlier revisions of DDR4.

The lowest I was ever able to go on my old kit was 16-16-16-32 CR1, but I had to pump the voltage way up to 1.48v. Not really worth it, and it's close to Intel's limit of 1.5v.

This kit however handles 14-14-14-34 CR1 with just 1.35v.

 

[Carfax83]

so it does
there is some interesting info here although they use fairly lose timings

Yeah I've seen that before. The reason why the Haswell variants have such low write performance compared to the other CPUs is because the uncore frequency is much lower.

Default uncore for Haswell-E is 3ghz, whilst Ivy-E is running it at the same speed as the core. Skylake has a default uncore of 4.1ghz..

This was the memory benchmark for my old kit using default timings but with the uncore at 4ghz. As you can see, the write bandwidth has increased tremendously compared to even running it at 3.5ghz.

The problem is though, is that with Haswell-E, combining high uncore frequency and high DDR4 speed can result in serious instability.. That's why I don't bother overclocking my uncore past 3.5ghz any more, as there is hardly any real world benefit, and you can have serious issues...

 

[Dasa2]

figured it was something like that
still interesting that higher frequency made no difference due to the memory controller\uncore limitation but the loser timings at higher frequency did create a small drop
so the uncore frequency is bottlenecking the write speed while having a moderate affect on read and a small affect on copy\latency

looking forward to seeing how tight you can get the timings
it will also be interesting when we get to see how broadwell-e handles ram

 

[Carfax83]

so the uncore frequency is bottlenecking the write speed while having a moderate affect on read and a small affect on copy\latency

Yep, as evidenced in my final tweaked timings. As you can see, read went up by 1GB/s, but copy had a large increase of 5GB/s.. Latency also dropped slightly as well, but not significantly.

Write performance remained the same because as I said earlier, it's linked to the uncore frequency..

This benchmark was run with the uncore at 3.5ghz.

it will also be interesting when we get to see how broadwell-e handles ram

Yeah I'm interested as well. I think Broadwell-E has a higher uncore frequency than Haswell-E, but not as high as Skylake-E.

Also, Broadwell-E's memory controller is likely tweaked to perform better, especially when it comes to writes which is Haswell-E's weakness it seems.

I'll probably pick up a Broadwell-E CPU myself if the opportunity presents itself. I didn't pay much for my 5930K, since I got it on sale at Microcenter for $389.

If I see a good sale for the 8 core variant, I'll buy it and sell my current chip. Or I might just wait for Skylake-E, especially if it will have PCI-e 4.0..

 

[Dasa2]

looks like they have improved write speed
http://www.overclock.net/t/1599068/6850k-vs-5820k

Currently using 33 uncore, can't go any higher on this board.

 

[Carfax83]

Yeah, that's a huge difference in write performance. My uncore is clocked 200MHz higher, and my RAM is at 3200MHz and he still whoops me.

See, that's why you should probably always wait for the revision :\