Sandybridge for overclocking, two solutions review.

[chinobino]

Multi-GPU scaling on MIVE

3DMARK03 (MIVE, 2600K D2, 5518Hz LS, 2x2G Kingston DDR3, Enermax 1250W)

Single 8800GTX - 45686

Two 8800GTX SLI thru NF200 - 67908

Two 8800GTX SLI by native SB lane split x8+x8 - 68582

So if you run 2-way it's wise to try avoiding NF200 bridge to reduce latency and have little better performance. All three tests with MIVE, coz UD7 don't allow have NF200 away if I recall correct

Full speed (x16 + x16) SLI hampered by the NF200!?!

[Edit: Asus web site reports; "Expansion Slots 4 x PCIe 2.0 x16 (single @x16, dual @x8, triple @x8, x16, x16 )"

So dual = x8 + x8 but triple = x8 + x16 + x16 ?

Am I missing something here? ]

I would love to see this re-tested with some more recent hardware (i.e. GTX480/580 or 5870/6970).

Damn you have a lot of tools!

Awesome write up

[dinos22]

good work TiN, i love your ghetto setups and unbelievable hardware setup and knowledge. It's awesome having people like you around that really like to open up the bonnet so to speak and go for broke

In terms of your summary and conclusions i want to respectfully disagree with you . I will tell you about facts that i base my opinion on below.

In terms of benchmarking UD7 is ahead in some things and behind in others. Most of the 3D benching will see equal scores other than 01 where UD7 is way ahead and possibly other system intense tests when pushing the system to the limit counts > ie hypers at monster timings and clocks.

Your 3DMARK01 difference seems a bit smaller than what people are experiencing, if you look around you will see up to 6K difference between the setups with UD7 being the faster one (4K difference link, Matose sees 6K difference GBT vs M4E). Ive had a quick test at your frequencies and UD7 scored way higher in system tests.

Have you done proper RAM testing comparisons and SuperPi efficiency with the boards? You'd expect to see ‘01 and ‘03 being run with 21.33 ratio with 7-7-6 on hypers as most guys would do when pushing hard. Have you tested 24.00 ratio with Asus board, does it work? It might be interesting to see some benchmarks with that 24.00 ratio but no one can get it to work online. I don’t have M4E but DLX board has Pi troubles with 21,33 ratio with ram i can run at 2200 7-7-6-20-1T on UD7 and it refuses to boot with 24.00 with auto timings even on P8P67 DLX. Gigabyte P67 boards are clearly faster/better in SuperPi efficiency(read the whole thread) than M4E and it seems to clock hypers more stable than M4E from Jody's findings on XS so far. One of your screenshots is showing 666Mhz ram clocks with cas9, what's up with that . That's the you test when you compare board bro , particularly old school benchmarks. You are a hardware guru though so i understand that these things maybe dont matter much in your testing

That 03 run you did, have you noticed GT2/3/4 that are heavy GPU tests are all the same between the platforms? You would expect NF200 to impact on those 3 tests, especially the last one no? NF200 does add latency I dont disagree as ive done testing on X58 platform but in that test it would be logical to see some of those being slower but they are equal.

UD7 has looser LLC implementation as seen which won't impact on benchmarks as we are forced to bench to frequency walls in Sandy Bridge. UD7 runs cooler than M4E as per your screenshots. UD7 is $100 cheaper as well. You could almost buy a UD3R for the price difference lol so if you get stuck in some tests where NF200 adds latency Bob's your uncle with UD3R.

Bios is still a bit early days on SB boards in general and UD7 is no exception. It is now starting to getting better with OC fail and other niggles people have with some of the latest bioses. LLC will get improved as well but yeah without it being a subzero setup its not as important as X58 so all those amazing extras like voltage monitoring on every single volt you can think of and PCie lanes control and bunch of other stuff is useless on Sandybridge

Too bad these platforms are restricted to just good watercooling or chilled water. They would be so much more fun if there was no MHz wall to really get the most out of the platforms.

It would be extremely cool if you do all top end GPU GTX580 hardware tests. I'd love to see how they stack up.

[computerpro3]

Just wanted to give a serious thank you for the review. Just outstanding work, and I will be purchasing a M4E as a result.

[TiN_]

Dinos, thanks

I have only single 480 here atm, maybe will test sli later more.

Ive had a quick test at your frequencies and UD7 scored way higher in system tests.

Post them?

I tried mem testing, but mine UD7 failed to work good with Hypers, sometimes even fail to POST on 4G and 800MHz dram clock with Dominators. With PSC no issues, same bios same settings.
Had no such issues with MIVE with same sticks, so I decided to skip that test right now, not to make false conclusions based only on exact system components. Later I will check.
And I dont have fancy 2400MHz DDR3 sticks to test 24.00 multi, sorry.

For 3D01 - that 4K difference was run with different memsettings a little I expect this is kind of software issue, and it may not be limited only to MIVE/UD7, maybe anybody compare another boards too in same environment?

I agree with you completely about BIOSes, too early for robust and flawless firmwares yet. But EFI looks much smoother, and mive have lots more ability to do fine tuning VRMs, clocks, etc.
Also I hate having automatic GBT BIOS backup by HPAing primary HDD, crushing RAID array if I forgot to disconnect it and reset CMOS. That's still here since P965-DQ6 i own years on. But i skip that part from review, it's not much relevant.
Normal guys should run SAS RAID cards, not teh ICH integrated madness .

LLC will get improved as well

Mmm, how? It's important always, not because of subzero but because it's about actual power delivery during transients, no matter of cooling. It's critical thing in overclocking, and margins even less than X58, coz 1155 CPU's much more sensetive to power stuff, it's not like GT on X58, pushing 1.9V .
PCI-E lanes switching is used on UD7 too, no exclusions (why otherwise GBT put NF200 there then?. And I can't see why it's useless, 1155 SB is quite limited by PCI-E lanes by its native.

Btw, any comments about NF200 cooling solution? Was not expected to see that

UD7 runs cooler than M4E as per your screenshots. UD7 is $100 cheaper as well.

I did stated that in conclusion, that UD7 runs cooler

UD7 have overvolts more with similar LLC settings, have more noise on plane, but VRM runs cooler in general.

As for price, it depends on retailer much. We had both boards (UD7 and MIVE) at about 350 USD here in Ukraine before they were pulled out by sata bug.

I like hardware, like messing with solder, that's obvious fact, I don't deny that. So review bit hardware-biased in a way, rather than testing all possible software things and tricks. I never even do copywaza or similar stupidity during benchmarks

[dinos22]

Why don't you setup a review page for extreme reviews like this as request hardware directly from manufacturers rather than have to resort through third party help. You can certainly setup a unique testing website and Im sure once you build it up a little all the companies that want to do some decent testing would send hardware that way.

Also I hate having automatic GBT BIOS backup by HPAing primary HDD, crushing RAID array if I forgot to disconnect it and reset CMOS. That's still here since P965-DQ6 i own years on. But i skip that part from review, it's not much relevant.
Normal guys should run SAS RAID cards, not teh ICH integrated madness .

what you need to do is copy main bios to backup bios, that stabilised the bios and stops bios backup tripping early

ALT+F12 command during startup will let you do that



Dominators, which ones, i have GT2000C7 and GTX2 they absolutely kick ass with this board. Maybe use F7e bios as thats where i really sort of tested a lot of superpi seriously

i was told that LLC can be improved, i dont know how as im not a bios engineer but if that is possible i will follow up and send you a new bios. I've had a different LLC bios sent for UD5 a while back so i know they can do some stuff with it for sure but to what extent and how i have no idea.

you misunderstood or i didnt word it properly, i was talking about powering off PCIe slots, things like that

[TiN_]

Why don't you setup a review page

I'm not quite want to do reviews 24/7 tbh. I like messing with hardware more than writing articles and making neat looking charts with useless fps or 3dmark numbers
I do run own website and own two servers for my internal R&D, but most effort i put to design own stuff and devices,rather that reviewing already done hw.
Just sometimes like to review something interesting, and I'm amazed about sandybridge, really best platform last few years, crushing LN2 records with air

As for bios - I'm not about reflashing bios from second chip, but about having GBT HPAing my HDD and breaking RAID by its bios "backup" stuff.

Corsair Dominator GT 1866C7

LLC can be improved,

It's a bit hardware stuff, not just pure BIOS.

Ah, as for PCIe, yea, switching I my mind mean switchswitch lanes I (and not only I, also guys who run multiGPU subzero) found similar feature useful on EVGA mb's. And it's really easy to do, so it's def not a con to have it)

[Raja@ASUS]

They'll have to revise in HW to change the LLC parameters to any appreciable degree Dino - BIOS can only change the applied VID. Voffset and the FB gain/compensation is all hardware set so maybe a board revision or something of that ilk.

[dinos22]

X58 is still THE platform for multi GPU subzero because of Mhz wall on SB hence why im telling its as useless as tits on a bull, you know what i mean dude

[TiN_]

Yea, but not everyone (i'd say almost no one in general) are packed with all that megabucks 990X and sets of 580's all under LN2, but lots of people had fun with 5+G aircooled/h20 sandy.
World doesn't stop on WR ranking on bot I found awesome to have 300USD CPU on Intel desktop board on air running faster than LN2-cooled to bones i7 950 placed in 400USD mb

[dinos22]

so are we discussing affordability now or benching 4way?

whoever can do 4xGPUs on LN2 surely fits the profile of someone benching a 980X or 990X

i agree about SB and that it's awesome fun however you can have just as much fun with a $150 P67A-UD3R board and most people will buy that not a $450 board LOL. M4E is the least affordable SB board on the market

Anyone that want to do decent benching will also have to buy a load of SB chips, Dumo is considering working for the local Microcenter after buying a miilion CPUs to find a 5.8GHz chip

Here is an 01 run man just did a quick one now with system tests, first go, didnt try anything special just ran dragos, lobbies and CH with 1.9LOD you said you used. I disabled SLI so i'm only using one GTX295 core, easily another 2K on top of what you are doing

[sin0822]

Hey Tin, just 1 thing other than the other stuff you wrote which i won't get into(you are always correcting me, so yea):

The chip you call a "frequency divider" is actually not. Sure in the past years a freuqnecy divider woudl have been teh most simple way to go, allowing both of the new PWM outputs to be in phase with the original, but technology has progressed. So it is labeled 617C, its actually an Intersil ISL6617 Proprietary Phase Doubler. it is designed to allow for the increase in phase count to 12 or 24 phases, specifically in conjunction with any of the ISL63XX. Including the ISL6366 which is on the UD7.


Now intersil did have many initial thoughts on how to design their phase doubler, YES they did think about using a frequency divider because that would allow for both outputs to be in phase with the pwm input, but that just wasn't good enough for them. They decided that they could further extend the function of the PWM, so they put a lot of work into a phase doubler, that would allow for each channel A and channel B to work out of phase. Channel A will take the full PWM frequency and then it will switch to channel B, so on and so forth. One channel would work in phase with the initial PWm input, and the other would swap when necessary to evenly spread the load. Of course they brought current balancing and offsets into play. The actually technical document they submitted to the US patent office was very long, anyways they did it and i can't explain it that well in common speak or i will get beaten up by you for it.


Honestly i thought it was a frequency doubler, as those are just as common as frequency dividers. but this sure does amaze me, and it probably might impress you by how intuitive they got. Although it is very complex and I have to agree a frequency divider would have been just much simpler, extending the PWM like they did is very promising. i personally cannot wait to see what Intersil will do with Digital PWMs, because it sure seems like they know what up.

[dinos22]

here's another one with low RAM speed

as you can see you are missing 2K in UD7 run and i'm using a weaker card (single core of 295 is slower than 285)

[TiN_]

Dino

as for buying load of SB chips,

Well , not everyone 980X even go pass 6G CPU-Z, no matter they cost triple more than top SB Lets skip pricing atm. Both UD7 and MIVE was sold at about same price locally, until bug was revealed. Right now we dont have neither boards on sale, so it's pointless to talk about virtual prices.

Nice score, i will do run today. 5166 is in my range.

I don't have 295. Mine gfx used was EVGA 285 with 2GB of vram. So it's single-gpu.

sin0822

I have full ISL6366 DS, managed to download it until ISL pull it off site, and read it thru carefully

Just two questions.

Channel A will take the full PWM frequency and then it will switch to channel B, so on and so forth

What happen to Channel A then when we put full PWM frequency, as you state, into channel B?
What frequency will we need on ISL6366 output to get 24-phases sequenced without overlap?

I checked ISL outputs very carefully, with analyzer of 350MHz bandwidth, and did not find any ..... ...... mm, signals which could double actual phase count. There are few methods, but this is not here.

[dinos22]

prices are still relevant and i am using current prices in Australian market. In fact UD7 price remained the same around $350 while M4E first launched at $500+ and now in mid-$400 so i am quoting current prices and our resellers definitely sell boards still.

[sin0822]

Dino



Well , not everyone 980X even go pass 6G CPU-Z, no matter they cost triple more than top SB Lets skip pricing atm. Both UD7 and MIVE was sold at about same price locally, until bug was revealed. Right now we dont have neither boards on sale, so it's pointless to talk about virtual prices.

Nice score, i will do run today. 5166 is in my range.

I don't have 295. Mine gfx used was EVGA 285 with 2GB of vram. So it's single-gpu.

sin0822

I have full ISL6366 DS, managed to download it until ISL pull it off site, and read it thru carefully

Just two questions.


What happen to Channel A then when we put full PWM frequency, as you state, into channel B?
What frequency will we need on ISL6366 output to get 24-phases sequenced without overlap?

I checked ISL outputs very carefully, with analyzer of 350MHz bandwidth, and did not find any ..... ...... mm, signals which could double actual phase count. There are few methods, but this is not here.

Well for your question i worded it a bit weird, but idk how else to w it so i gave you a snippet of the patent(actually the into). My information on how it actually works is not from the ISl6366 datasheet, yea no need that datasheet can be found on google on alldatasheets they DLed it like you and I before.

I found that it has nothing to actually do with the ISl6366 itself as much as the doubler is designed to work with the current sensing and PWM outputs of the ISL6366. I thought that segment from the patent would explain it to you better. its not part of the ISl6366 datasheet.

maybe this can explain it a bit better for you??? i can agree with you that it is a bit much to have 24 phases, but its really isn't bad at all, it evenly spreads load and increase efficiency.

Tobe honest i do not think they want to make how it works so public, which makes more work for you and me, me having to ask my friends who know a patent attorney.

Anyways tin nice job, i just wanted to point that out.

[TiN_]

Dinos, I'd love to, but all I have is 9800GTX+, monster 285 and single 480. Well 8600GT too, but thats not interesting

Sin, Nice patent stuff. The point is that it doesn't occur in UD7 I had stready waveforms w/o switching. Check out youtube vid

[sin0822]

I am just showing you what it says they are using for their phase doubling tech. I saw your tests and I am not saying they aren't right, i am saying maybe they are hiding their tech, you saw the correct delays tho. IDK maybe you need more channels. a VRTT test will show, if a UD4 has MUCH better transient response than a UD7, then yes maybe you are right, it is expected that it will have better, but significantly better, like something like 30% then yes divider. Of course they patent this, not divider tech. They also don't hide every datasheet on ISl6617 if its just a divider, as that is an accepted method.

but hey i just wanted to point that out. That patent is a few years old BTW.

EDIT last time i tried your youtube video it didn't work.

[TiN_]

So it's again not much hardware issue, but rather software/bios things, etc, as be both run UD7.

[sin0822]

i run a UD7 too lol. BTW it gives me some amazing bench scores.

Anyways Tin, listen, i think when it switches it might just turn off the phase, like what GB advertises as phase switching tech comes straight out of some things that Intersil writes. Sometimes when i had a stupid little DMM on the choke, it would lose the voltage. of course just like it has modes of operation when higher loads occur it might switch its operational mode. you did your readings of an MLCC cap, would you be able to see a phase or two drop off like that? I believe what you posted as results from your scope are true.

BTW why can't GB change LLC? i have seen it vary BIOS to BIOS just from user feedback in some of the UD7 or UD4/UD3R threads i go on, i have seen people complain about one BIOS then have it fixed in another, LLC issues.

BTW since these boards the UD7 is being replaced with the UD7-B3, i wonder what they have changed, as they changed some things on the UD4-B3 right?

[dinos22]

try a newer bios like F7e, if it doesn't give you the scores you may have been reviewing a faulty board

[TiN_]

Sometimes when i had a stupid little DMM on the choke, it would lose the voltage.

That's called just bad probe connection, m8

dinos22
It works I can try F8x you used.

[sin0822]

That's called just bad probe connection, m8

hahahaha lol, i soldered a wire to the chokes leg to make life easy

[TiN_]

hahahaha lol, i soldered a wire to the chokes leg to make life easy

Then bad soldering. All chockes connected to single power plane there, they are physically interconnected, if voltage drops - then your rig is down. so check your DMM/probes/connections etc.

It's offtopic btw

Dinos

maybe u have 9800GTX+ or 480 to try ?

[sin0822]

Then bad soldering. All chockes connected to single power plane there, they are physically interconnected, if voltage drops - then your rig is down. so check your DMM/probes/connections etc.

It's offtopic btw

Dinos

maybe u have 9800GTX+ or 480 to try ?

lol what else do you want me to talk about. lol. yea ill try to make it happen again, i swear it just let off. Anyways sorry to bring this stuff up tonight while Dino posted, the thing is that i just found out finally what it does, and I was planning to post when i was certain. And that was tonight. I understand you are probably overwhelmed by Dino's really great argument and my un needed one. You just make the UD7's VR which is proven through the UD9 look like its just stupid and you made me question it.

[chew*]

I have an M4E on the way eventually direct from ASUS. I also have a UD7, direct from GIGA.

Although my time is severely limited I will pit both boards against each other at balls to the walls settings in a large range of benches.

I would do it anyway for my own personal data of what boards are best at what so I don't have to second guess crap when I tackle a specific bench.

Usually I keep this data for personal use only however.......

I will not be going over electricals on the boards nor will I pretend to know what each chips purpose is or how it works. Sin0822 I like your reviews but just a friendly suggestion, no information is better than misinformation.

I also don't play favorites nor will I ever, credibility comes first for me and I will not ruin mine by sucking up to any manufacturer.

To drive that point home I would do another compare of ASUS PRO versus UD4, however the ASUS PRO has memory compatibility issues for balls to the walls benching as well as general stability issue if not really cold and the UD4 failed on my bench table and can no longer accept "bench" vcore, it's usefull as a daily driver now and thats about it when it decides to boot.

Until then guys seriously this little back and forth stuff between ASUS and Giga is getting kind of old.

Both have hit or miss products in the product cycle, thats life.

If I lose my connection with either or both manufacturers for this post......so be it.