New Memory Highs: Corsair and OCZ Introduce DDR550
by Wesley Fink on February 19, 2004 6:03 PM EST- Posted in
- Memory
Why Would Anyone Want DDR550?
When computer users visit some of the hard-core Enthusiast sites on the web, one of the first things that they will likely see is memory overclocking. You will see dropped multipliers and incredible FSB settings, all in the effort to raise the memory speed to the highest level possible. Skeptics always ask, why all the fuss? What do you really gain?3.2GHz CPU Speed, Different FSB and Memory Speed
To try to answer that question, we took a close look at our 3.2ES. The unlocked CPU allows multipliers to be selected from 12 to 16, which means with DDR550 memory, we should be able to compare 3.2GHz performance at two very different settings. At a multiplier of 12 and a bus setting of 266 (DDR533), we can achieve 3.2GHz. This can also be achieved at 16x200 (DDR400) with a default 3.2 Pentium 4. So how does performance compare at the same 3.2GHz?| 3.2GHz Performance - Different FSB - Corsair XMS4400 | |||||||
| Memory Speed & Multiplier | Memory Timings & Voltage | Quake3 fps | Sandra UNBuffered | Sandra Standard Buffered | Super PI 2M places (time in sec) |
Halo DX9 1024x768 | Comanche 4 1024x768 |
| 400DDR 16x200 |
2-3-3-7 2.55V |
378.5 | INT 2624 FLT 2680 |
INT 4528 FLT 4527 |
108 | 54.9 | 60.25 |
| 533DDR 12x266 |
2.5-4-4-7 2.75V |
411.9 | INT 3345 FLT 3362 |
INT 5842 FLT 5819 |
102 | 55.2 | 63.25 |
| % Change | +8.8 | +26.5 | +28.8 | -5.6 | +0.6 | +5.0 | |
At the exact same CPU speed, lowering the multiplier and increasing the bus setting and memory speed increases performance 26% to 28% in Memory. This translates to improvements in gaming performance from only 0.6% in Video-Card/CPU bound games like Halo to 8.8% in Quake 3. Keep in mind that the CPU speed is exactly the same, only the bus speed has been increased. This is what the overclockers are aiming for, at even more extreme levels than we have tested here.
In this example, there is an ongoing argument as to which really contributes to the increased performance. Some believe that the increase in the FSB to the CPU is mainly responsible for the improved performance on Intel 875/865 boards. They argue that a high FSB combined with slower memory at aggressive timings will give the exact same results. We decided to test that argument with the fastest DDR400-433 memory we have compared to this Corsair XMS 4400 at DDR533.
3.2GHz CPU Speed and Same 533 FSB, 1:1 Memory vs. 5:4 Memory
To test this claim, benchmarks were run at the same 3.2GHz CPU speed achieved with the same 12x533 multipliers. This means that the FSB was a constant 1066 in these tests. Memory was tested at DDR533 with slower timings and compared to another memory at DDR 426 (5:4) at the fastest timings available of 2-2-2-6.| 3.2GHz & 1066 FSB - 1:1 Memory vs. 5:4 Memory | |||||||
| Memory Speed & Multiplier | Memory Timings & Voltage | Quake3 fps | Sandra UNBuffered | Sandra Standard Buffered | Super PI 2M places (time in sec) |
Halo DX9 1024x768 | Comanche 4 1024x768 |
| 426DDR 5:4 12x266 |
2-2-2-6 2.65V |
412.3 | INT 3061 FLT 3116 |
INT 5678 FLT 5695 |
101 | 55.35 | 63.49 |
| 533DDR 12x266 |
2.5-4-4-7 2.75V |
411.9 | INT 3345 FLT 3362 |
INT 5842 FLT 5819 |
102 | 55.2 | 63.25 |
| % Change | 0 | +8.6 | +2.5 | -1.0 | -0.3 | -0.4 | |
The results for 1:1 with slower timings versus 5:4 with fastest timings are very interesting. While synthetic memory tests like Sandra show 1:1 memory 2.5% to 8.6% faster, all the games perform essentially the same to slightly faster with the slower memory at 5:4 with faster timings. This pretty much destroys the arguments made on some sites that memory timings do not matter in high-speed memory. If Memory timings did not matter, then DDR423 would certainly perform much slower than DDR533, when in fact DDR423 is the same to faster.
In fairness, no one would argue that DDR550 at 2-2-2-6 is no faster than DDR400 at 2-2-2-6. All things being equal, faster speed will be faster. With today's memory, however, all things are not equal, and we are often forced to use slower timings as a trade-off for faster 1:1 memory speeds. Sometimes the trade-off is not worth it, and you are better off with slower memory running at 5:4 with the fastest timings possible. Do you save money with this approach? Not really, since the best DDR400 to DDR433 memory that will actually run at 2-2-2-6 timings is about as expensive as the faster memory with more relaxed timings. If you already own very fast DDR400 to 433 memory, however, 5:4 may be just as good an option as high speed memory with slower timings.
Facebook
Twitter
RSS
13 Comments
View All Comments
klah - Saturday, August 14, 2004 - link
good articlePumpkinierre - Friday, February 20, 2004 - link
#9 Perhaps you're right but all the reviews I read on the IC7 associated the game accelerator with PAT eghttp://www.lostcircuits.com/motherboard/abit_ic7/6...
the speed increase is of the same order as PAT ie 2-5% eg:
http://www.xbitlabs.com/articles/mainboards/displa...
But I acknowledge I made an assumption so could be wrong.
#11 I get a ~1% increase in performance on 3Dmark2001 and 3dmark2003 (default settings benchmark and catalyst 9800pro drivers) when runnung [email protected] at mem. 5:4 2237 (OCZ 2x256Mb PC3200 platinum original SS) Game Accel.-auto. compared to 2.8 1:1 mem. oçlocked 2237@216MHz GA-F1(memory wont handle Street Racer). In general game play, the 1:1 feels smoother in my opinion.
TrogdorJW - Friday, February 20, 2004 - link
"Let me tell you , it does make a speed difference. Everything benches 1-3% higher."So it's a measurable difference, but not noticeable. :p
I'd still like to see real benchmarks on a variety of applications rather than just take someone's unsubstantiated claim that 3.2 or 3.4 GHz with PAT is best. With the system bus OC that you get, I don't think buying a 3.2 would be faster than buying even a 2.8C and OC'ing to 3.2. However, I have neither so I have no way of knowing.
retrospooty - Friday, February 20, 2004 - link
BTW , on my Epox 4pc3a+ I can enable, or disable PAT at any speed or ram ratio.Let me tell you , it does make a speed difference. Everything benches 1-3% higher
retrospooty - Friday, February 20, 2004 - link
PAT works on i875 fulltime... Even at 5:4Those settings you are referrring to " F1 or street racer" are not actually PAT , that is GAT, its just Abit's memory tweaks. On the I865 Abit boards, GAT can enable PAT (much to Intels displeasure) but on I875 PAT is always enabled. the GAT settings are just memory tweaks.
Pumpkinierre - Friday, February 20, 2004 - link
Turn it off and see if it makes any difference. On Abit you cant run F1 or street racer PAT settings above CAS2 or on 5:4. I dont know about Turbo but it doesnt add all that much anyway.Icewind - Friday, February 20, 2004 - link
Uhhh, you wanna bet? Im running PAT at 5:4 ratio with my Corsair 3700XMS on my P4C800-E Deluxe bud at 3-4-4-8.Pumpkinierre - Friday, February 20, 2004 - link
#4 and #5 you cant run PAT with 5:4 ratio nor with memory timed higher than CAS 2. So, Trog, your best cpu is a 3.2 or 3.4. That way you get the high speed with small overclock allowing the use of DDR433 low latency BH5 chipped memory (o'clocks to 450). Amongst the cheappies, the 2.8 is the best and use OCZ 466 gold which holds CAS2 up to 420 and then 2.5 through to DDR500. Or else use the 3500 Mushkin or OCZ low latency.#5 I dont see why you cant get 2225 at ddr500. Those graphics cards have got 2.2ns chips and run at DDR700-1000. I'm not sure about the latency but if you lower the speed you can improve on the latency. I'm waiting and I'll buy when it comes out. All these DDR533 and 550 seem to be a rehash of the same thing and missing the low latency quality that is required for PAT to be enabled.
Icewind - Thursday, February 19, 2004 - link
Condsidering its the ASUS's highest end board, they probably left the PAT on, and why you would want it off in the first place is beyond me.Considering the limations of the current breed of DDR chips, a 2-2-2-5 or close to that at DDR500 simply isn't plausible from a manufacturing/cost point.
DDR2 aint looking much better either, it runs at 4-4-4-12 settings stock. So I think the days of low timing memory are going to be going the way of the do do.
TrogdorJW - Thursday, February 19, 2004 - link
I'd be curious to see a roundup of various benchmarks done with the varying memory speeds and timings, sort of like what you started with by comparing 3.2 GHz at 266 MHz bus to 3.2 GHz at 200 MHz bus. Here's what I'm thinking:Get a 2.4C, 2.6C, 2.8C, 3.0C, and 3.2C. (Or use your P4 3.2ES, I suppose.) Then do a variety of benchmarks (i.e. not just Quake 3, SuperPi and Sandra) at reasonable settings and memory timings.
From what you've shown in this article, a 2.4C overclocked to 3.2 GHz will outperform a 3.2C at stock bus speeds. However, it could do this with low latency 5:4 ratio or higher latency 1:1 ratio settings. What appears to be the best choice? A 2.4C would require a 266 bus to reach 3.2 GHz, where a 2.6C would "only" require a 246 MHz bus, a 2.8C would require a 229 MHz bus, and a 3.0C would need a 213 MHz bus. At those bus speeds, the 2.4 and 2.6 overclocks would need to use either lower timings or a 5:4 ratio, but the 2.8 and 3.0 could probably get 2-2-2-6 (or 2-3-3-7 timings?) with a 1:1 ratio. Also, how does PAT affect things? Did you have it enabled on the 1:1 OC and not on the 5:4 OC? I wan't sure.
I know, it's a lot of work and may not be that useful to many. Still, it's something I would enjoy seeing when/if you get the time. Anyway, with 2.4C, 2.6C, and 2.8C all costing about the same amount right now, I'm not really sure where the best buy is. 2.4C would probably be fastest if you could get really expensive RAM to handle the high OC and the processor worked at 3.2 GHz, but do 2.6C or 2.8C overclock higher on average, due to the higher multipliers?
Great article, though. Nice to see that low timings can still match higher clock speeds. Also, any chance of seeing a similar roundup using Athlon XP? (I would also say Athlon 64, but they don't seem to have enough OC headroom.) Athlon XP might not be able to make use of anything over DDR533 - or maybe even DDR500 - but I haven't seen any good comparisons on it.