Once and for all-Which DDR2 for a 1:1 ratio?

[ianx]

I hate to sound stupid but...kick me, I'm stupid!
I do not understand, yet it should be simple, how to match DDR2 to a Intel FSB.
My logic tells me that if the FSB is 1066, ie on a Q6600, then I should buy
DDR2 1066/PC2-8500 memory. However I am confused about the difference
between the memory clock and the I/O bus clock. If 1066 is 266 quad pumped,
and everyone (so they claim) is running 333 to get a 3GHz clock on their Q6600,
then which memory is 1:1 at stock FSB and what do you want at 333?
Again, I feel stupid asking this but as I don't have Mr. Shimpi's or Laporte's
personal phone number, and I see so much drivel on this subject when I search
other sites/forums, I am putting my faith here.

 

[Goldfish4209]

Well, the FSB for a Q6600 does run at 1066Mhz, but the cores run at 266x9, making 2.39999...Ghz. (Rounded up to 2.4) To get 1:1, I would just get DDR2-800 and overclock to 1066. This should be doable on most quality RAM kits. Once you get past 800Mhz, prices rise dramatically, and it's just worth it more to get the slower kits and overclock.

 

[GarfieldtheCat]

Originally posted by: Goldfish4209
Well, the FSB for a Q6600 does run at 1066Mhz, but the cores run at 266x9, making 2.39999...Ghz. (Rounded up to 2.4) To get 1:1, I would just get DDR2-800 and overclock to 1066. This should be doable on most quality RAM kits. Once you get past 800Mhz, prices rise dramatically, and it's just worth it more to get the slower kits and overclock.

++

1:1 ratio at 400FSB = 3.6GHz CPU

Unless you think you are going to do some hardcore OC-ing of your Q6600, DDR2-800 is all you need.

 

[ianx]

Assuming the chart below is correct (please advise if not) and if I understand
what GarfieldtheCat is saying, then (a): the FSB of the CPU and the I/O Bus (memory)
need to be the same to attain a 1:1 ratio, (b): a stock Q6600 should be used with DDR2-533,
and (c): a Q6600 OC to 3Ghz (9x333) would then need DDR2-667 to maintain a 1:1 ratio.
If this is all true (and please god make it so, my head hurts) then WTF is 800 and 1066 used for?
I only ask the last question out of curiosity, I assume (nay, pray) the answer I was looking for
is that I need DDR2-667 to run my new (yet to be completed) Q6600 box at 3GHz with a proper 1:1 ratio.

Standard Name Memory Clock I/O Bus clock Module name
DDR2-400 100 MHz 200 MHz PC2-3200
DDR2-533 133 MHz 266 MHz PC2-4200
DDR2-667 166 MHz 333 MHz PC2-5300
DDR2-800 200 MHz 400 MHz PC2-6400
DDR2-1066 266 MHz 533 MHz PC2-8500

I'm sorry, it seems my chart will not display as typed.
The columns merge together.

 

[Quiksilver]

1:1 with DDR2-533 (266*9 - 2.4Ghz [Stock])
1:1 with DDR2-667 (333*8 - 2.7Ghz, 333*9 - 3Ghz)
1:1 with DDR2-800 (400*6, 2.4Ghz [Stock]) 400*7 - 2.8Ghz, 400*8 - 3.2Ghz , 400*9 - 3.6Ghz)
1:1 with DDR2-1066 (533*5 - 2.7Ghz, 533*6 - 3.2Ghz, 533*7 - 3.7Ghz, 533*8 - 4.3Ghz, 533*9 - 4.8Ghz)


If my math is right (and rounded correctly) that little chart I made should be right, also I left out the lower multipliers because the numbers would be below stock.
I don't even know if these are achievable so their bolded

If you want 3Ghz and nothing higher, go with DDR2-667 (Of course some memory modules could be overclocked to DDR2-800 speeds). If you won't mind running a little more than 3Ghz go with DDR2-800; at least with DDR2-800 you can OC a bit more and keep your 1:1 ratio when you machine seems to be bogged down in the future or you just feel the need to ramp up the power.

 

[brawleyman]

Is there any performance issues using a ratio of 2:1 with memory? I have an AMD system where the HT bus runs at 200mhz. PC2 800 would run at 400mhz. I have heard talk about using 1:1 ratio of ram to bus speed, but was just wondering about odd ratios like what I mentioned. Would I be better off getting PC2 400 (200mhz 1:1) with low latencies or PC2 800 (400 mhz 2:1) with higher latencies?

 

[BonzaiDuck]

There is no hard and fast rule about this.

You can under-clock DDR2-800 and DDR2-1000 modules to run at lower speeds in a 1:1 ratio.

It may even be prudent to purchase higher-rated modules to run at lower speeds if they are "latency-elastic" and run at nice, tight latencies. Often, only the performance reviews will show this, and the stock latency settings reported in the maufacturer specs won't indicate it. Sometimes, it is a matter of making an "educated-gamble."

You could also find DDR2-667 modules that run at tight latencies and can be over-clocked to the same latencies or just a tad looser to higher speed in Mhz.

It isn't the rating of the RAM per se. Run a search of reviews that show how tight the latencies can be set for various modules at various speeds. You might luck out and get DDR2-667s that can run higher at lower latencies, and several D9-Micron-based modules of DDR2-800 and DDR2-1000 are capable of running at DDR2-667 with very tight latencies.

What you want to look at are the bandwidth benchmarks that results from modules -- either way.

For people who feel a surge of enthusiasm about over-clocking, but who only look at FSB over-clocking or running RAM at high-speed specs on a divider, the latency issue is vastly over-looked. It's tedious, and you'd want to do some reading, but it pays off.

 

[BonzaiDuck]

FN:

If Q6600 is pushed to 3 Ghz at 1,333 FSB, then nominal RAM speed at 1:1 ratio is 667. But whatever RAM you'd buy, you'd be better to get RAM that will run at 667 with latencies 3,3,3,6 than at 4,4,4,12. Some here have shown that you can set some Crucial modules to 3,3,3,3 -- defying conventional DDR1 wisdom.

 

[nefariouscaine]

Originally posted by: BonzaiDuck
FN:

If Q6600 is pushed to 3 Ghz at 1,333 FSB, then nominal RAM speed at 1:1 ratio is 667. But whatever RAM you'd buy, you'd be better to get RAM that will run at 667 with latencies 3,3,3,6 than at 4,4,4,12. Some here have shown that you can set some Crucial modules to 3,3,3,3 -- defying conventional DDR1 wisdom.

i had a long reply to all this and canceled it out on accident...

but i'm running those modules at the same settings as before (3-3-3-3 T 1 Trc 7) but lowered the voltage to 2.01v

i'm unsure of your insistence on 1:1 ratios as stock clocks for 1066 fsb and DDR2-800 is on a 2:3 ratio

this also implies you can run your fsb at 1600 (400 quad pumped) drop your cpu multi down to 8 and be hitting 3.2ghz (generally also not too hard to do...) run the ram 1:1 and do DDR2-800

but factors other than ram need to be mentioned over all as to what motherboard to use as not all are quadcore OC friendly. Are you outright pinging on getting a G0 stepping quad? Is the power supply a reliable & stable unit.

if all you are worried about running is 1333 fsb at 1:1 settings and nothing else just find good quality D9 based memory but be sure your mobo has options for up to 2.2v on the ram as this is a generalized manufacturer spec requirement (not always needed as is my case)

Crucial Ballistix is a good bet, and most vendors have their share of other kits based on quality IC's too but the trend is Crucial always has D9's. Cost hasn't been mentioned nor has the quantity of ram you're looking for. I'm sure you have enough food for thought - I'll look for more posting on this

 

[BonzaiDuck]

Nefarious --

Just an FN.

I've finally tested these DDR2 1000-s at a divider, loosened timings and higher speed.

For instance, 5:6, DDR2-800. Note that I've done these tests -- not for hours of stability, but just to see what the bandwidth benchies show. I set them at 4,4,4,8, 1T, but didn't drop tRC -- (which would have a noticeable pos impact on bandwidth.)

It tentatively seems to be a wash. I can run multiplier 9 @ approx. 3.2 Ghz or 350 CPU_FSB, and the memory at 800 (I THINK that's 7:8). The bandwidth benchies show about 9,200.

Now -- if I pursued this avenue, I think I could tighten tRCD down to 3, but I don't think I could get tRAS lower. With the harmless lowering of tRC, I might get somewhere between 9,500 and 9,700. Not anticipating any big improvement on 1:1 over-clocking, though.

AigoMorla ran these at something like 4,3,4,8 or 9 and 1000Mhz, with the voltage set at 2.2. He says it was too much for them. I don't think the lower latencies are much stress, not as much as running them full-bore WITH the latencies and upper-end voltage.

Overall, as I'm tentatively implying, it seems to be a wash with the best setting you can get at lower DDR speeds.

I think they'd have to go with CAS / tCL = 4 when they push these above DDR2-740. That was my experience, anyway.

But I still need to see how -- and WHY -- (for myself) you can lower the tRAS to your setting. with tRC set to 7, that's a significant improvement, and you don't need to do more than a mild over-clock, as we've discussed here and elsewhere.

Anyway, even the Crucials are pretty cheap right now, and as you've stated, their limited-lifetime warranty coverage is reliable.

 

[BonzaiDuck]

Another FOOTNOTE to ALL:

Here's a page out of a June 2007 Tom's Hardware performance / benchtest review of several DDR(x) modules -- including some DDR2's:

Everest bench results

Note the Corsair CAS 3 DDR2-800 modules, which apparently run at DDR2-800 with 3,4,3,9 settings. The write-read-copy results from Everest on these makes them appear to be SLUGS comparied to my Crucials running at DDR2-667 through DDR2-704 (with different voltages, but same timings and tweaks, or 3,3,3,6,1T tRC=9).

Back again later -- I'm being nagged to chauffeur someone to an appointment .. .

 

[BonzaiDuck]

Continuing "the Footnote:"

The THG article suggests that "latency timings are not as important for DDR2 as they were for DDR(1) . . . "

Yet, with their Corsairs at DDR2-800, timings 3,4,3,9 -- no indication about command-rate or bank-cycle-time -- their Everest results totally suck compared to my Crucials at lower FSB with slightly tighter timings. That is, the Corsairs' tRCD is 4 versus my 3; their tRAS is 9 compared to my 6 -- and we don't know what THG did with respect to tRC -- if anything at all.

Just a minute -- I'll post my Everest bench-result screenie . . . .

 

[BonzaiDuck]

Everest Bench Result for Crucial DDR2-1000s 2x1GB @ (DDR)= 704 Mhz, timings 3,3,3,6,1T,tRC=9

The VDIMM voltage required for this setting was 2.2V -- exactly, or the recommended maximum.

The same settings can be made at (DDR)=667 Mhz volted at 2.13V. Another member has run similar modules at even lower timings and at this speed (as I understand from his posts and PM's) -- with voltages lower than this.

So the essential Everest bench results shown in the screenie are:

Memory Read: 9,753 MB/s
Memory Write: 6,396 MB/s
Memory Copy: 6,581 MB/s
Latency: 51.8 ns

and you can compare those to the Corsairs in the THG article.

 

[ianx]

Well...after all this CAS/RAS tech talk, I think what BonzaiDuck mentioned in his 12/12 posts was the info I was looking for.
Buy faster DDR2 and run it slower, ie DDR2-800 at 333MHz to obtain lower (faster) latencies.
That way I can maintain a 1:1 ratio with the FSB and the memory clock. I thought that's what AnandTech has always preached.
As for THG, f**k them, I trust no one but AT. On another note, I see a near flame war going on (another thread) on the merits of OCZ modules.
WTF, AT seems to like these guys, I personally run a pair of OCZ4002048PFDC-K and have no problems. Should I stay away from
these guys in my new build?

 

[BlueAcolyte]

Don't be so easily influenced, but I haven't heard too many good things about OCZ either. Maybe it's because people only complain/post when they have a problem.

 

[GarfieldtheCat]

I've used several sets of OCZ RAM without problem. Every company will have some bad RAM sticks, from the best to the worst memory company. Also, lots of times, the problem may not be the RAM itself, but the treatment of the RAM (overvolted, OC'd too much, etc...) or that the desired OC wsan't attained.

I've also used Corsair, Mushkin, and Patriot, and all seem to be about equal to OCZ. (Haven't had any RAM failures yet <knocks on wood>)

 

[BonzaiDuck]

Ianx --

There's still always some element of "crap-shoot" in this.

While I've built all my systems and several others since maybe 1995, I only started seriously over-clocking around early 2004. Maybe I already told the story. OCZ Gold EL DDR-500's were "the hot item." But at best, you could only tighten them from 3,4,4,8 to about 2.5, 4,4,7. No matter how close to DDR-400 you set them, you couldn't use any tighter timings, and they were no match a pair of Mushkin 2,2,2,5 DDR-433's. But there were also these OCZ Platinum DDR-400's, supposedly made with Aeneon chips, that had stock timings of 2,3,2,5, and you could run them up to DDR-452 at just under the recommended voltage without loosening the timings one bit.

So when I bought my DDR2-1000 Ballistix, it was still a gamble. Everyone on the forums that has them, or even ddr2-800 / or -1000 Tracers, says "Gee!! Aren't they wun-dur-full!?"

AigoMorla says he was running his Tracers at the 1000Mhz clock, (something like) 4,3,4,8, at 2.2V, and now they're candidate for "RMA." Some others here have been more daring at 2.3 or 2.4V. Aigo says he was doing a lot of 24/7 "folding@home." I think he said something like "D9 . . . . not all it's cracked up to be . . . . " But do we know of anything better?

I"m just guessing that running them at tight timings, @ between 667 (2.125V) and 704 (2.2V) is safer. So -- guessing -- I suppose I'm "hoping" . . . . .

All I can say: shop around, look all over the world for reviews. The rest of the insight, I hope you find here.

 

[The-Noid]

Those are very good results bonzai, congrats on the high read speed.

680i is a very good memory clocker @ CR1.

 

[BonzaiDuck]

Originally posted by: Yoxxy
Those are very good results bonzai, congrats on the high read speed.

680i is a very good memory clocker @ CR1.

Now, see . . . . That's something I wouldn't have foreseen at all when I bought the Striker.

I'm a little puzzled at the moment, though. I'm using v 4.00.976 of Everest Ultimate. What I just noticed, probably wouldn't be just a software bug in Everest, because CPU-Z also picks it up.

Occasionally, with the 3,3,3,6,1T timings, after bringing the system out of "Standby," both utilities show the latencies as 2,3,3,6. There aren't any errors in operation. It could also be something having to do with my TrackMania game, because I've seen this happen after coming out of STandby, then playing the game, and then raising Everest or CPU-Z.

Notwithstanding, I found an intermediate point between my 3.0 Ghz setting and the 3.17 setting @ 352 Mhz -- something close to 3.06 @ 343 Mhz, with the same timings and command-rate. I must have run PRIME95 Blend for 10 hours today, and the small-FFTs test for 8 hours -- no errors, no warnings.

Again -- puzzled by this "2,3,3,6" phenomenon . . . . nothing in the event-logs to indicate any malfunction . .