Revisiting 680i vs 975X

[lopri]

Kinda busy with personal life and been from somewhat of a distance to AT. Can't spend much time toying around with hardware nowadays but managed to flash my semi-abandoned EVGA 680i board to the latest P26 BIOS - and the results were refreshing. I will update the comparisons as much as I can but for now I have a few screenshots of stability testing as well Super Pi runs.

Orthos
SP 32M 1T: E6400 @3600MHz (8x450), 800MHz / 3-3-3-1T
SP 32M 2T: E6400 @3600MHz (8x450), 900MHz / 4-3-3-2T

1T - 3DMark01
2T - 3DMark01

1T - 3DMark06
2T - 3DMark06

1T - SP 1M
2T - SP 1M

I've run more benches but those will have to wait. But the core observations are:

1. With 2.30V-ish Vdimm, 800MHz / 3-3-3-1T is quite easy. Depending on DIMMs, 900MHz / 3-4-3-1T with 2.40V is also possible and totally stable. The performance is, of course, FANTASTIC.

2. As you can see, the benefit of 1T is big enough to cover the disadvantage of lower memory speed vs FSB. This trend has continued throughout various testings. This means, 1T lets you over come,

• 2-1. 1:1 is NOT the rule of thumb on Intel platform anymore. In previous AT memory reviews, there used to be a comment regarding disadvantage of running memory async to FSB. But with the emergence of 1T on 680i platform, we might see a different evaluation methodology in the future.
  
  2-2. Memory running slower than FSB is not a penalty any more! In my testing, FSB was 450, and memory was running @400MHz (DDR2-800). This used to be a no-no on Intel platform because CPU was fed data via FSB. However, in order to match the performance of 9:8 (FSB:Memory) / 1T with 2T timing, the DIMMs needed to be clocked at least to 1100MHz /4-4-3 which required more voltages. This reminds us once again that latency and bandwidth are relative to each other.

3. According to my (limited) testing, the performance of 680i with 1T (which is now quite easy and stable) finally surpasses that of 975X, clock for clock. Of course this is with 975X with 2T, but currently there is no means to set 1T on 975X so the comparison should be considered valid. This became all possible thanks to P26 BIOS which also majorly fixed the SATA issue on my board.

(..to be continued)

 

[lopri]

Everest V4.00

1. 680i w/ E6400: 3600MHz (8x450), Async memory (slower than FSB) DDR2-800 @3-3-3-8-1T
2. 680i w/ E6400: 3600MHz (8x450), Sync memory (1:1 with FSB) DDR2-900 @4-3-3-9-2T
3. 975X w/ E6600: 3600MHz (9x400), Sync memory (1:1 with FSB) DDR2-800 @3-3-3-8-2T
4. Opteron 165 @3000MHz (9x333)
5. Opteron 165 @2800MHz (9x311)

Now enjoy the views.

Memory Read
Memory Write
Memory Copy
Memory Latency
CPU Integer
CPU Multimedia
CPU Compression
CPU Floating 32bit
CPU Floating 64bit
CPU Floating 80bit

Update: 4GB with 1T (4/19/07)

Did a quick test with all 4 slots filled with 1GB DIMMs. No problem running 8x400 with 1T @4-3-3 timing. The highest it can do is below:

http://img353.imageshack.us/my.php?image=4gb1tiz3.png


Update: (4/22/07) PCMark05 Test Suites

975X with E6600 @3.60GHz, No RAID

680i with E6400@3.60GHz, No RAID

Update (4/24/07) Cinebench 975X vs 680i

It seems like the extra L2 cache of E6600 (vs E6400) is big on this bench in that even DDR2-900 / 1T on 680i isn't able to overcome the deficit.

Cinebench: E6600 @3.60GHz on 975X

Cinebench: E6400 @3.60GHz on 680i

Final update in this thread (4/28/07)

3DMark05 with 8800 GTS 320 19608 :Q

3DMark06 with 8800 GTS 320 11491 :shocked:

GTS 320 was clocked @636MHz, which effectively gives 648MHz performance. (memory = 999MHz)

Orthos under Vista @3.60GHz / 1.52V (actual)

http://img291.imageshack.us/my.php?image=e6400b1fb6.png
http://img234.imageshack.us/my.php?image=e6400b2vw2.png

(Shameless plug: My Q6600 has arrived from MWave and the E6400 used throughout this thread is out for sale here.)

 

[lopri]

Added 3DMark scores.

 

[aigomorla]

GOD i hate my eVGA board now.

The thing has so much bugs, and it errors and craps out after 4-5 months according to a lot of people over at xsforums.


EVen with the P26 bios, this board still has massive instability issues, overheating SPP problems.

Save yourself the headache, if your not goign to touch SLI ever, go with the BX2. Such a more stable, and mature chipset vs. the 680i


If you really want a 680i board, i suggest you wait for dfi to come out with there 680LT edition. That board looks hwot.

lopri can you do a 1m also? i dont know how to compare your results to my 32m, more like im too lazy to let my comp sit there while it pulls a 32m. i know there great at testing stability, but lets see what the differences are between 1m on 1T and 2T.

Great post btw...

 

[lopri]

The SPP and MCP both get very hot. I'm assuming that you're using the 'mandatory' fan that comes with the board? MCP is capable of reaching 80C and SPP.. like 100C without active cooling. I just got rid of stock cooling and did my own.

Antec 900 & EVGA 1
Antec 900 & EVGA 2

Added Super Pi 1M digit results.

P.S. Oh and make sure you set the HyperTransport multi to X3 (or even try X2 if you're trying 2000+FSB) It definitely has an effect on stability as well as 1T. Also I observed that some sticks like blue slots better and others like black slots, so try different slots also. I learned that MCP (southbridge) needs a good amount of voltage, not only the SPP (northbridge). I set SPP @1.45V and MCP @1.60~1.65V depending on the number of HDDs and stuff.

 

[aigomorla]

yeah i already did that. The board is very limited compared to my brothers BX2 which has a E6300 @ 3.4ghz


Hehe.. and you dun know me very well i see:

http://i125.photobucket.com/albums/p73/aigomorla/IMG_0574.jpg

http://i125.photobucket.com/albums/p73/aigomorla/FinalTemp.jpg

Temps arent the issue to me, its the stupid board.

TY for the 1m... now i can see if its realyl worth going 1T

 

[lopri]

I see it. I could never build a water rig like that.. (you know, it's intimidating. ) Do you have any active cooling around the VRM area?

yeah i already did that. The board is very limited compared to my brothers BX2 which has a E6300 @ 3.4ghz

What do you mean by 'limited'?

Also if you haven't already, try x3 multi for HT link. Many weird issues disappeared with that.

 

[aigomorla]

Originally posted by: lopri
I see it. I could never build a water rig like that.. (you know, it's intimidating. ) Do you have any active cooling around the VRM area?

yeah i already did that. The board is very limited compared to my brothers BX2 which has a E6300 @ 3.4ghz

What do you mean by 'limited'?

Also if you haven't already, try x3 multi for HT link. Many weird issues disappeared with that.

meaning my board is still bugged, and im too lazy to rma it. I cant pass 400 fsb more then 2. The board wont go any higher.

And active cooling?
http://i125.photobucket.com/albums/p73/aigomorla/IMG_0552.jpg

http://i125.photobucket.com/albums/p73/aigomorla/IMG_0553.jpg

I had a post on it eariler. Its the best thing for the 680i board. You just need to make sure your mounts will clear it. And yeah it helps stability a lot when you have active cooling on it.

Make sure you pick up some cheap 1.99 DDR ram sinks with it. The full cover one, and steal the thermal tape off them. The thermal tape alone costs you 2.99 for a simular piece, so i dont understand of paying more just for the tape, when you can get more and ram sinks for cheaper!

 

[lopri]

I ran some more 'synthetic' benches: the latest Everest (ver. 4.0)

This time, there is another contender in this small comparison - Opteron 165. There are total 5 different configurations to compare and they are:

1. 680i w/ E6400: 3600MHz (8x450), Async memory (slower than FSB) DDR2-800 @3-3-3-8-1T
2. 680i w/ E6400: 3600MHz (8x450), Sync memory (1:1 with FSB) DDR2-900 @4-3-3-9-2T
3. 975X w/ E6600: 3600MHz (9x400), Sync memory (1:1 with FSB) DDR2-800 @3-3-3-8-2T
4. Opteron 165 @3000MHz (9x333)
5. Opteron 165 @2800MHz (9x311)

Now enjoy the views.

Memory Read
Memory Write
Memory Copy
Memory Latency
CPU Integer
CPU Multimedia
CPU Compression
CPU Floating 32bit
CPU Floating 64bit
CPU Floating 80bit

680i with 1T even gives an edge to 2MB L2 over 4MB L2 on 975X in many cases.

 

[lopri]

OH and one more thing: The new P27 BIOS finally stabilized 500FSB on my board. (Original revision with C55.. or NF 590 SLI MCP) But as predicted, there isn't really much to say about it performance-wise. I've never really understood the virtue of sky-high FSB on Intel platform. (likewise lowering CPU multipliers. Why people do that? ) Core 2 Duo's multipliers are already low enough! Up'ing FSB doesn't just NOT raise the performance but also stresses the board/memory. I tested @3500MHz with x7 and x8 (500FSB and 437.5FSB, respectively) and the performance was in general lower with 7x500 combo.

 

[terentenet]

Originally posted by: lopri
OH and one more thing: The new P27 BIOS finally stabilized 500FSB on my board. (Original revision with C55.. or NF 590 SLI MCP) But as predicted, there isn't really much to say about it performance-wise. I've never really understood the virtue of sky-high FSB on Intel platform. (likewise lowering CPU multipliers. Why people do that? ) Core 2 Duo's multipliers are already low enough! Up'ing FSB doesn't just NOT raise the performance but also stresses the board/memory. I tested @3500MHz with x7 and x8 (500FSB and 437.5FSB, respectively) and the performance was in general lower with 7x500 combo.

Core 2 Duo may not need stellar FSB speeds, but quad cores do in order to have larger CPU-RAM bandwidth and keep all four cores busy without waiting for FSB clearance to request more info from RAM.
It was proven back when it was first released that Kentsfield was bottlenecked by the low 1066 FSB speed. I'd say 1333 is minimum, 1600 or more is recommended.

 

[yacoub]

How does 650i fit into this picture? Expect similar results as 680i?
650i does not support EPP but at DDR2-800 it's not like RAM is using EPP for its timings unless it's really crappy RAM that can't get decent timings at DDR2-800.

IOW, would 650i with DDR2-800 @ 1T w/ a 1066MHz C2D would be better than
650i with DDR2-1066 @ 2T w/ 1066MHz C2D?

 

[defiantsf]

Thanks Lopri! Good info to have as I build my upcoming E6600 rig with a planned upgrade to Q6600 later in the year.

 

[lopri]

Originally posted by: terentenet
Core 2 Duo may not need stellar FSB speeds, but quad cores do in order to have larger CPU-RAM bandwidth and keep all four cores busy without waiting for FSB clearance to request more info from RAM.
It was proven back when it was first released that Kentsfield was bottlenecked by the low 1066 FSB speed. I'd say 1333 is minimum, 1600 or more is recommended.

Gotta agree with you there.

Originally posted by: yacoub
How does 650i fit into this picture? Expect similar results as 680i?
650i does not support EPP but at DDR2-800 it's not like RAM is using EPP for its timings unless it's really crappy RAM that can't get decent timings at DDR2-800.

IOW, would 650i with DDR2-800 @ 1T w/ a 1066MHz C2D would be better than
650i with DDR2-1066 @ 2T w/ 1066MHz C2D?

The results should be, under the assumption that there is no artificial limit in BIOS, identical or nearly identical in that 1T will pull out superior CPU/memory throughput than any other chipset available currently. The most ideal configuration would be (obviously) 1T with 1:1 (FSB:Memory) or even higher memory speed with 1T but right now 1T seems only stable up to 820~830MHz. (which translates to 400~415FSB, not too shabby on its own)

But if I remember correctly there is a voltage limit on 650i Ultra boards. (vDIMM<=2.10V?) That'd be the biggest hurdle for 1T operation. You may want to keep an eye on different manufacturer's board specs. For 800MHz/3-3-3-1T, you'll need 2.20~2.30V for memory.

 

[lopri]

For instance check this review out for 650i chipset.

http://www.vr-zone.com/?i=4874

So it's safe to say that NV doesn't want 650i to overclock as well as 680i (for obvious reasons) but various vendors might figure out their own implementation. (They even let you SLI @8x/8x configuration!) All that I've read indicates that 650i SPP is nothing but the same 680i SPP with some function removed and/or binned. So in that sense it could be a luck of draw. But the same can be said for 680i also.

 

[lopri]

Update: 4GB with 1T

Did a quick test with all 4 slots filled with 1GB DIMMs. No problem running 8x400 with 1T @4-3-3 timing. The highest it can do is below:

http://img353.imageshack.us/my.php?image=4gb1tiz3.png

Everest V4.00 seems to report values from various sensors on the board correctly. However, it's still software and it can't detect the lies the sensors tell it. For example, from my DMM measuring vDIMM on my board is 0.07~0.08V higher than what the sensor says. So in this case, the actual vDIMM would be more like 2.40V and I don't feel comfortable giving more than that with 4 sticks of RAM. Also at this point what'll crap out first is the NB, not the memory IMO. 4 sticks have been always more stressful to the memory controllers than 2 sticks.

 

[lopri]

Update: (4/22/07) PCMark05 Test Suites

975X with E6600 @3.60GHz, No RAID

680i with E6400@3.60GHz, No RAID

Amazin, isn't it? A few months ago 680i couldn't even come close to 975X in this specific benchmark. Look how things have changed.

 

[lopri]

Update (4/24/07) Cinebench

It seems like the extra L2 cache of E6600 (vs E6400) is big on this bench in that even DDR2-900 / 1T on 680i isn't able to overcome the deficit.

Cinebench: E6600 @3.60GHz on 975X

Cinebench: E6400 @3.60GHz on 680i

 

[Shimmishim]

how about ddr2-1200 2T on the 680i?

i'm sure that should overcome the deficit!

 

[o2brew]

What voltage are you giving the RAM to run at 1T? And are you still running the Crucial 10th Anniversary sticks on that board? I'd like to get mine running at 1T, but would like some idea of where to start. I'd be happy even with 4-4-4 1T if that would take less volts.

 

[lopri]

o2brew: It'll really depend on the type/quality of modules but my guess is 2.2~2.3V for 800MHz/3-3-3-1T for non-defective Micron D9 sticks. For 4-4-4-1T.. I haven't really thought about it but will check it out. By logic it should take less volts but you never know with this kind of things.

Shimm: 1200MHz takes too much volts for my liking so I'm trying not to go there. (especially since I want to stabilize things with 4GB)

 

[lopri]

So my interim conclusion with regard to squeezing the best performance out of Core 2 Duo platform is:

1. 400~450-ish FSB under 1066 strap (unless you own a $1,000 CPU that gives you the freedom of multipliers as well as a board that lets you choose strap)
2. 1:1 memory with CAS 3 and Command Rate 1T (I believe 1T will soon come to P965 as well as the upcoming Bearlake chipset)

This means;

1. 
   1. Forget about 500~FSB. Not only the performance isn't guaranteed (due to the transition to 1333 - or even 1600 - strap), but also the chipset will be under extreme stress which will affect stability as well as its life-span.
   
   2. No need to push memory MHz higher with loose timings. Granted most 975X and P965 still don't have an option of 1T, but even then tight timing/low MHz vs loose timing/high MHz difference is often times slim-to-none.
   
   3. Push the FSB while trying to stay under 1066 strap. If your board has a certain FSB area where it has hard timing booting or maintaining stability (A.K.A. FSB Black Hole), chances are that's where the strap change occurs. It varies board by board so you need to experiment on your board. In general, on 975X it's somewhere between 420~450, 420~430 on 680i, and 380~420 on P965.
   
   4. If you can stabilize 1T on your board (680i and upcoming P965s), 800~850MHz (1:1) /3-3-3-1T will give you the best performance than any other configuration. Don't be a victim of memory makers' advertising strategy (World's first DDR2-1249390000 with Mega Ultra Turbo Xtreme Cooling - patent pending!)
   
   5. I'd still reserve my final opinion on this matter, but if you can't keep clocking 1:1, just raise the FSB but leave the memory around 800MHz/3-3-3-1T. As my (again, limited) testing shows, even with the memory running slower than FSB, the tight timing/latency can make up the loss from not running memory:FSB=1:1.

 

[theprodigalrebel]

Nice to see the coolest platypus on AT back in action!

 

[FireChicken]

Nice getting 4gigs to run at 1T. I read some where that 4gigs wouldnt run at 1T but guess that was wrong. I couldn't get my 4gigs to run at 1T at any ram speed and they are D9s. these sticks were tested at OCZ and are not defective. Wonder what the trick is

 

[lopri]

Final update in this thread.

3DMark05 with 8800 GTS 320 19608 :Q

3DMark06 with 8800 GTS 320 11491 :shocked:

Orthos under Vista @3.60GHz / 1.52V (actual)

http://img291.imageshack.us/my.php?image=e6400b1fb6.png
http://img234.imageshack.us/my.php?image=e6400b2vw2.png

(Shameless plug: My Q6600 has arrived from MWave and the E6400 used throughout this thread is out for sale here.)