Lowest-Latency DDR3?

[TIGR]

Priority 1: CAS/CL (in nanoseconds, not clocks!)
Priority 2: frequency (chip data rate)

Is there anything better, in terms of the above priorities, than G.SKILL F3-16000CL6D-4GBPIS? According to their specs, they'll do 6ns CL. The closest competitors I know of are the F3-17600CL7D-4GBPIS, F3-17600CL7Q-8GBPIS, and F3-17066CL7D-4GBPIS, all of which come out to ~6.4ns. They are higher-frequency kits but latency is my first priority.

Coming in at 7.0ns, there's also the F3-18400CL8D-4GBPIS, and from Mushkin, the 996829 and 996902 kits.

I mention these other ones because they're the closest possibilities I know of but I don't know how well they overclock compared to the F3-16000CL6D-4GBPIS. Should I assume the F3-16000CL6D-4GBPIS are best for my priorities, or is there a product here or one that I don't know of that can outdo them?

This is for a testing run.

 

[mastrdrver]

For benching it kind of depends on the app. Though if your doing this with your i3, you're going to be limited a little by the on chip qpi link from one die to the other. If with your AMD, latency is all that matters. Though past 1333 c7 with about 2600 cpu-nb tangible results are pretty null unless your using a memory intense application like Adobe PS or something similar.

Also if I'm not mistaken your only talking about access latency. Who cares about that? I'd be more concerned about how long it take the complete the entire transaction from access to precharge.

To answer your question though, you'd be better off with some hyper sticks cause I think they are still the best thing available at the moment. Ewiz has Super Talent sticks that are 2000 c8 for 150 and 2200 c8 for 180. Wait around a few days and they will run some promo that will end up being 15 bucks off.

 

[TIGR]

For benching it kind of depends on the app. Though if your doing this with your i3, you're going to be limited a little by the on chip qpi link from one die to the other. If with your AMD, latency is all that matters. Though past 1333 c7 with about 2600 cpu-nb tangible results are pretty null unless your using a memory intense application like Adobe PS or something similar.

Also if I'm not mistaken your only talking about access latency. Who cares about that? I'd be more concerned about how long it take the complete the entire transaction from access to precharge.

To answer your question though, you'd be better off with some hyper sticks cause I think they are still the best thing available at the moment. Ewiz has Super Talent sticks that are 2000 c8 for 150 and 2200 c8 for 180. Wait around a few days and they will run some promo that will end up being 15 bucks off.

What depends on the app? As I said, this is for a testing run (not personal use). I will be testing the impact of latency on performance in a number of areas, from gaming to workstation use. Testing will include AM3, 1156, and 1366-socket platforms with CPUs from Phenom 550 through 1090T and Core i3 530 through i7 980X.

Despite being the last in order in memory transactions, CL has a well-earned place at the front of how timings are generally expressed (CL-tRCD-tRP-tRAS). Because memory is often accessed sequentially (from the same row), CL has the most significant and universal impact on memory performance. As tRAS (the delay between the precharge and activation of a row), tRP (time required to terminate one row access and begin the next one), and tRCD (the delay between row and column activation) all deal with row changes, their impacts on performance are significantly less consistent and consequential.

That said, it would be great to get some 6-6-6-18 DDR3-2000+, so if you know of any—at any price—I'm all ears. Even better would be some CL7 DDR3-2300+ and I'd love to get my hands on some DDR3-2500 that could do CL8. I have some Patriot DDR3-2500 but it is 9-11-9-27 IIRC and tighter timings aren't happening on it.

All the Super talent CL8 DDR3 I know of is 8-8-8-24, vs the G.SKILL model I noted at 6-9-6-24. The Super Talent CL8 DDR3-2000 you suggested has significantly higher latency than the G.SKILL CL6 DDR3-2000, which goes against the first priority I typed in the OP. CL8 @ DDR3-2200 is closer, but at 7.3ns CL, is easily beaten by the G.SKILL at 6.0ns. Super Talent does have some 7-7-7-21 DDR3-2000 (Project X) which wins out in tRCD and tRAS, but still suffers from higher CL.

Also, fixed the link in the OP.

Anyway, I'd be interested in any of the following:

DDR3-2000+ that can do better than 6-9-6-24
DDR3-2300+ @ CL7
DDR3-2500+ @ CL8

 

[TIGR]

Bump!

 

[mastrdrver]

Hyper period.

For AMD, 1600 6-6-6-18 will beat that G.Skill stuff at those timings. Performance on AMD is in tRas. 8-8-8-24 and 7-7-7-24 at same speeds will see no difference in memory benches outside of allowable error. AMD doesn't care about cas. Real difference is seen in tRas and to an extent tRcd.

For 1366 your not going to be running three sticks at cas 7, 2200+ and not run slower than cas 8, 2200.

For 1156, Hyper will do 6-6-6-18, 2200+. Nothing else really will. The cheaper bbse and power chips will do timings like those G.Skills but will be slower in your benches.

There is a reason sticks with hypers cost so much. They are the best right now. Though they are going away which is probably why your seeing them so cheap. They're just being dumped on the market to get ready for the next thing.

Do some reading here and I think you'll get what I'm trying to say.

With the old Intel fsb yes, cas timing did matter most. Though with ddr3 and Intel now joining the on die memory controller club, the other 3 "main" timings are starting to effect performance.

Are you planning on doing any benching with 4gb sticks? I see you mentioned workstations and I've been watching those kits with 4gb sticks slowly decline in price. By the end of the year things could get interesting. Just not seeing any reviews on them or any performance kits with them.

 

[Helper]

mastr explained it very well. Look at timings instead of NS if you need performance. As NS is latency it shows up in memory latency benches but in real world usage such as gaming the lower CAS you have, the faster it becomes.

Pi's are notting... they use cheap and slow chips with high tRCD and tRAS like mastr said.

CORSAIR DOMINATOR-GT 4GB (2 x 2GB) 240-Pin DDR3 SD...

They'll do 7-7-7-21 at speeds over 2000 to a degree. 2150-2250? Dominator GTs and HyperX T1s used to be the best known DDR3 sticks a year ago they cost 400-500$. Now you can get them for half the price.

 

[mastrdrver]

Well until Hypers went cas 8, 2200 everything was cheap then that was 2000+.

I still think the Super Talent 2200 cas 8 is the cheapest option right now on Ewiz. Everything else is 200+. Though those low volt 1866 Kingstons look interesting suppose to be binned Hypers. Of the few pics I've seen, they seem to scale well until about 1.5-1.6v.

 

[LAN_deRf_HA]

6-6-5-18 @ 1600 Mhz http://www.mushkin.com/Memory/Redline/998730.aspx Good luck finding it in stock though.

 

[mastrdrver]

Guess I should have just linked this huh? Doh!, I'm feeling a little slow this evening.

 

[TIGR]

Sorry to not reply for a while, forgot about this topic.

mastrdrver, I now have a set of the G.Skill G.SKILL F3-16000CL6D-4GBPIS on the way (those are the 6-9-6-24 @ 2000 ones you said would be beaten by 6-6-6-18 @ 1600 RAM). Been looking for some 6-6-6-18 @ 1600 RAM so I can run a quick side-by-side to test your assertion, but all I've come up with are Corsair CMG4GX3M2A1600C6 and TW3X1600C6GT, which I can't actually find for sale so far. (OCZ OCZ3FX16002GK [EOL] is listed as 6-6-6-18 at Newegg but as 6-6-6-24 on OCZ's product page.) You said that Hypers will do 6-6-6-18 @ 2200+. What specific modules/kit can I test this out with?

I will indeed be testing 4GB modules as well as 2GB and 1GB modules.

 

[LAN_deRf_HA]

I'm pretty confident these sticks would do 6-6-6-18 @ 1600, maybe even lower. https://shop.corsair.com/store/item_view.aspx?id=766411
I'm tempted to get some because good ram like that is quickly going extinct and I'd love to have the insane flexibility of speed and timing combinations but it will really rape the bank.... on the flip side since the chips are reaching EOL maybe they'll retain value.

 

[TIGR]

Thanks LAN_deRf_HA, I'm going to hold out for a bit and see which modules mastrdrver was referring to that can do 6-6-6-18 @ 2200+. That would allow me to do some thorough impact-of-latency testing, starting at 6-6-6-18 @ 2200 and working my way down to see which changes have what impact. It would be nice to be able to do all the testing on one kit, so that other variables can be eliminated.

 

[mastrdrver]

Sorry if I came off sounding like a jerk initially. Now reading back over what I wrote I could see it come across like that. Just is case.

Now to the question: Super Talent CL8 2000 Super Talent CL8 2200

There is also the low voltage Kingstons, 1866mhz ones I think, that are suppose to be Hypers from pictures posted. Those two Corsairs kits you listed should be Hypers too. Basically just about anything 8-8-8-x 2200 or 2000 will more than likely be Hypers.

I don't think 6-6-6-18 @ 2200 is going to happen on anything short of a 1156 Lynnfield system with a very good IMC chip and matched dimms. Did you happen to see the link to the other thread on TPU I posted earlier? Eva did make 8-8-8-24 1T with a 980X at almost 2400Mhz. He also did 7-7-7-21 1T at just over 2000Mhz. Knock that down to two dimms on a 1156 and there may be hope for 6-6-6-18 at 2200.

I'd personally just go PM him here or over on i4memory.com. Most of what I'm commenting about is just from reading of his work over on his forum. He'd be the one you'd really want to talk to as he knows a lot about ram clocking on 1366 systems. AMD systems are pretty much what I've picked up over on XS and in personal testing.

AMDs don't care about bandwidth as much as timings. Speed at the same timings will probably see a slight increase in performance, but real change happens when you can move from 7-7-7-21 to 6-6-6-18 at the same speed. I even went 8-8-8-24 to 7-7-7-24 with all other timings/speed the same and saw a total of zero improvement in memory benches. No joke. I was stunned. Other benches didn't change either. Even more evidence was when I got some CL7 1600 Doms to put in the same Gigabyte board as some Ripjaws. The Doms at 1666 7-7-7-21 1T show same results in memory benches as the Ripjaws at 7-7-7-21 1T at 1333.

Please, please post a link once you get done with it all (maybe Christmas ).

 

[TIGR]

No prob at all mastrdrver, I prefer opinions that conflict with mine to ones that agree with me, because they make me think.

What motherboards for each platform (AM3, 1156, and 1366) do you think would be best for getting the absolute most out of a high-end memory kit for benching? Four PCI-Express x16 slots supporting SLI and Crossfire would be best as well. I have a 980x, 875K, and 1090T ready to roll, but am not so sure about the motherboards I have on hand. And, in my shoes, if choosing one kit to test the greatest range of latencies and clock speeds possible, which memory kit would you get?

I wasn't planning on running any synthetic benches—the real world ones are what matter to me and to my clients (mostly)—but it seems a shame not to, at least for the highest-end tests. Trying to decide whether or not it's worth the added time and hassle when I'll already have my hands beyond full.

I suppose now's the time to pick up a few blocks and pumps and go find a good car radiator or two. And run some new wiring in the shop.

 

[mastrdrver]

No prob at all mastrdrver, I prefer opinions that conflict with mine to ones that agree with me, because they make me think.

Ah, we think alike then.

On the board thing I'm not sure. I know currently no other AM3 board is more refined bios wise than the Crosshair Formula IV.

On all the other questions I would highly suggest talking to Eva2000. I'm not sure if anyone else has tested so many different memory kits than he has.

I've just done a lot of reading and testing only the couple of kits I've had in my possession. Eva has had way more than anyone should sanely have. I think he has some kind of ocd with memory kits, it a good way though.

I don't think I have enough experience with the different ram kits to send you to the right. Rather go to the master, if your will, than someone listening in the crowd.

Btw you going to probably find this link useful when testing what ever ram you get: Memtest86+ and what errors are related to