266 v. 333 DDR. The real increase?

[Bob151]

I've been lurking arround here and the OC forum for a week or so and I remain curious about something and I haven't found the explicit thread or article on Anadtech to reveal it to me. If its out in the open in a current thread, please forgive.

Q. All other things being equal, is there a significant performance increase when DDR memory speed goes from 266 to 333? My calculation indicate ~ 25% increase by the Mhz alone. Is it that way in practice, in the resulting benchmark of system speed? My most Mhz intensive appls will be 3D FPS games, not photo.

Thanks all.

 

[Soulkeeper]

depends on what kinda setup you have i mean if your running it in an athlon system
with a 133fsb then the extra memory throughput won't really make much difference
put in the P4 systems where the fsb is 533 and the chip needs as much bandwidth as you can get it does help.
if you are gonna overclock ddr333 is definately a bonus
but as far as real world performance goes when nothing else but the mem mhz is increased you won't really notice the diff unless the fsb is already eating up all the bandwidth needed
there are also other factors to consider such as onboard graphics that might share the memory and other devices that might access the memory too
personally i would build a system with ddr333 just cause i would prob oc the fsb right off the bat

well hope this helps
also read some of the articles over at tomshardware
cause they have recently posted some benchmarks where they have oc'd systems and compared them to mulitiple setups to answer questions like this

--Soul_keeper

 

[THUGSROOK]

heres a thread in the CPU/OC forum.

 

[rogue1979]

Thugsrook, those were some excellent, informative benchmarks. Lately, I have been sitting around trying to justify buying some Samsung PC2700 and maybe a new motherboard if needed. But, quite frankly, after looking at your results it seems that we are all getting caught up in 3DMark2001 and Sandra 2002 results. I don't know about some other applications, but it looks like gaming uses all the memory bandwidth it needs at 266MHz DDR speeds. If we already have a DDR system, looks like the only benefit from moving to 333-400MHz DDR is synthetic benchmarks. So for some fool like me looking for more gaming performance, a new motherboard and RAM is a foolish upgrade, better to use the money for a faster video card. With your P4 set up you adjusted this with memory ratios, on an Athlon system using the actual fsb speed to raise the DDR speeds (lowering the multiplier of course) would this make more of a difference in gaming performance?

 

[Bob151]

Originally posted by: THUGSROOK
heres a thread in the CPU/OC forum.

Thanks Thugsrook,

In your first message of that thread the columns of benchmarks look like the additional DDR bandwidth doesn't add substancially to the 3d game benchmarks, but does in Sandra and 3DMark2001.

But later you say, "to answer your question - the loss of memory bandwidth kills game benchmarks even up as high as 2.6ghz/145fsb."

How is this?

Thanks

 

[oldfart]

I disagree. I have much better game benches with higher DDR speeds. Here is a post from [ H ] forums. My own testing has very similar results.

Zroc over @ [ H ] forums did some testing on various oclocking combos. You can see the full read on the hardforum page.. I pasted it here for the AT guys to see:
..................................................................................................
Ok, I started out exploring memory timings with an 8IRXP, 1.8 OC'd to 2.4, a 512MB stick of Samsung PC2700, and a Ti 4600...if you care to review, check it here:
http://www.hardforum.com/showthread.php?threadid=386912

I wanted to explore mem timings a little more, so I grabbed a 4BDA2+ and a 1.6 chip, 256MB Samsung PC2700, and a Ti 4600, just to mix it up a bit. I must say, OC'ing on this 4BDA2+ is a freakin' breeze, very nice, though I like the layout, look, and features on the 8IRXP much better.

Anyway, to the point...on the 8IRXP, I basically found that using max memory timings at default speeds (1.8GHz, 266 mem) produced noticable gains - again, check that thread for the actual differences. But what puzzled me is that, at 2.4GHz, 354 mem, using max memory timings did practically nothing...1 point gains in Sandra, and actual tiny decreases in Quake III and 3DMark 2K1SE scores.

So, onto the 4BDA2+:
Seems like I got a really nice board, chip, and memory. The chip is a Costa Rica, pack date 4/17/02, which I thought would be crappy due to being Costa Rica, but right now, I'm running it a 2.4GHz/400 mem no prob, at default 1.5v...I'm stoked about that. The mem is also at 400 2/5/2/2, with still default voltage for the board (2.65). I fired up Prime95 on it when I went to bed, and it was still chuggin' this morning, so kick ass. This is all on a 300w SH power supply, too

Ok, benches are Sandra Bandwidth, Quake III 1.17 default fastest settings with sound, and 3D Mark 2001SE default.
Running at 2.4GHz, 400 mem, with mem timings of 2/5/2/2:
Sandra- 3058/3054 QIII- 298.6 3DMark- 11707

Running at 2.4GHz, 400 mem, mem timings 2.5/6/3/3:
Sandra- 3057/3055 QIII- 296.7 3DMark- 11684

Really similar to the 8IRXP/1.8@2.4/354 mem speed. Basically, using max mem timings at high memory speeds does practically squat. One point again in Sandra, a 1.9 frame increase is QIII (a .6% increase), and 23 points in 3DMark (.1% increase). At 400MHz memory speed, max timings gave me a 0 to .6% increase, where at default 266MHz mem speed, I got roughly 20 points outta Sandra, a 3.7% gain in QIII, and a 2.7% gain in 3DMark...not noticable, mind you, but enough to make me use those max timings (I had calculated that in the other thread to be worth about 3MHz of FSB for Sandra and 5MHz of FSB for QIII/3DMark 2k1).

Ok, let me REALLY drive this home...I set the timings back to 2.5/6/3/3, went into the BIOS, and added one measly point to the FSB, so it's now 2.41MHz and 402 mem speed:
Sandra- 3078/3074 QIII- 300.6 3DMark- 11768

There you have it...a direct comparison at high memory speed (400+) between max memory timings and MHz. I got much better gains from simply pushing the FSB up 1 more MHz than from using max memory timings. Guys and gals, I feel this is important information to the P4/DDR overclocking community. Yes, max memory timings make a nice little difference at 266 (or even 333, as I showed in a previous thread with an Athlon/7VRXP KT333). But for some reason, once the memory speeds get up there (350/400+), max mem timings seem to amount to s**t, and certainly don't appear to be worth sacrificing even one MHz of FSB overclocking.

Please, share your thoughts, post comparisons, or offer your theories, as I think this is important.

I'm gonna go have a little more fun with this 1.6, as I seem to have a gem...I'll post back with an update on it's progress

-------------------------------------

Update on the chip: up to 2.6GHz now, though I had to lower the mem speed ratio ta 2.0 versus 2.66.
I found this interesting, as the question gets asked a lot. At 2.6GHz/163(652)FSB/ 326 mem speed, I got:
Sandra- 2493/2488 QIII- 296.1 3DMark- 11760

In essence, the QIII and 3DMark scores are roughly the same for 2.4GHz/400MHz mem and 2.6GHz/326 mem. So it would seem that any of you that could run at 2.4/400, but lowered your memory to get anywhere between 2.4 and 2.6 should probably put 'em back

Gonna see how much more this chip'll do...

 

[THUGSROOK]

Bob151 by that i meant - 333ddr 4:5 is alot better then 266ddr 1:1.
its the same thing that oldfart is trying to say.

roughly it works like this.
2400mhz 133fsb 400ddr equals 2484mhz 138fsb 345ddr in 3dmark2000/2001se.
2610mhz 145fsb 290ddr couldnt even get close.

 

[Dually]

I don't think you would see that big on an increase at this point. It would only be noticed with file servers. Eventually graphics cards could use the room. AMD's increase of the l2 cache 256KB to 512KB will be noticed a lot more.

 

[oldfart]

I just ran a couple of quick Q3 demo_four benches. This is DDR300 Vs DDR400, but it shows if mem BW matters.

1.6A @ 2.4 GHz 150 FSB 1:1 DDR300
Q3 ver 1.30 demo_four
Normal..............................285
HQ....................................282
1024 x 768 x max quality..257

1.6A @ 2.4 GHz 150 FSB 3:4 DDR400
Q3 ver 1.30 demo_four
Normal...............................317
HQ.....................................312
1024 x 768 x max quality..278

 

[rogue1979]

oldfart, now I am getting really confused. It looks like the DDR speed does make a significant difference in gaming.

Anyone else have some results to show?

 

[brittlenet]

It really depends on whether or not you are video card or CPU limited. If you have a GF4 4600, most current games can't stress out the video card and are CPU limited. (The Unreal Tournament from 1999 is an excellent example). Obviously, if you are CPU limited, system memory bandwith will be much more of an issue.

That said, most future games are likely to add features that will be able to stress out a high-end graphics card, and will scale back the quality to accomodate lower end cards. CPU usage in the future will be largely dominated by AI, collision detection, etc, which is not as scalable, and must cater to the greatest common denominator of CPU speed.

Thus, game makers are likely to produce games that will run on mainstream (read: slow) CPUs, but will add features to take advantage of the best graphics cards (esp since eye candy sells). Given an extra $50 to spend on the CPU or $50 to spend on the graphics card, I would put that money into the graphics card until you get past the sweet spot. Then I would dump it into CPU performance.

If I may liberally interpret the benchmarks, Thugsrooks's results were based on a GF3 Ti500 while Oldfart's were based on a GF4 Ti4600 (correct me if I'm wrong, Oldfart, I saw one in your system specs and assumed you used it). Therefore, Oldfart's benchmarks were more likely to be CPU bound than Thugsrook's.

 

[THUGSROOK]

brittlenet very well put. i must also point out that my test was done on a SiS645 w/a 1.8A cpu.

if oldfart had a 4:5 ratio he could overclock to 160fsb 400DDR - then hed be a really happy oldfart

its a real shame that all i845/SiS645 motherboards dont have all the ratios available at all fsb speeds. it would make overclocking even more interesting.

here are a few ratios available from different boards...
3:6, 3:5, 4:6, 3:4, 4:5, 1:1, 4:3

 

[tenoc]

Hmmmm. How many of us can notice a difference between 257 and 278 fps?

I think I can tell the difference between Pesi and Coke, but not that!

 

[Bob151]

Thanks all for you contributions.

 

[Mingon]

The one thing people have not mentioned so far (i think) is that increasing memory speeds only help when they match or are less than FSB speeds / throughput. A P4 as mentioned will show good gains (at same settings) as memory speed increase as it is always running with less bandwidth than the FSB.

 

[THUGSROOK]

tenoc do you drive? do you speed? can you tell the difference between 55mph and 65mph? are you actually getting there faster?

thats what i thought

 

[Bob151]

Originally posted by: Mingon
The one thing people have not mentioned so far (i think) is that increasing memory speeds only help when they match or are less than FSB speeds / throughput. A P4 as mentioned will show good gains (at same settings) as memory speed increase as it is always running with less bandwidth than the FSB.

So the trick is to get at 1.6A or 1.8A? What is the FSB speed of those? Are those the 533s?

 

[Mingon]

The p4 1.6 and 1.8 are all 400mhz fsb based processors. The 'trick' with proccessors is they can only currently use as much memory bandwidth as the fsb allows. So the closer you get to synchronous bus speeds the better.

An ideal solution for p4 is dual channel ddr running at 266mhz while the fsb is running at 533mhz, which is what intel have planned 'sometime' soon. Its a shame they were too awkward to give nvdia a liscense for the P4 as the nforce chipset would have been excellent.