[INQ] Performance RAM will damage your Nehalem

[leomax]

AT Core i7 - Is High VDimm really a Problem?

[Charles Wirth]

Good article by Gary.

[T_Flight]

I liked the way he put it. He said "The rules have changed completely for Intel"

I also liked the hints that there are things we do not know, and it depends on the BIOS options. To me that means study and follow directions lest I let the magic smoke out and will be taking another trip to the bank. I'm real clear on this. I'm NOT going above that vDimm limit. I'm starting with RAM that is in spec and if I ever do go above it, it will be after I have a complete understanding of what I am doing.

I also know that this is not a 775 based system with a FSB and a MCH. It is a Core i7 1366 CPU with an IMC and is a completely different technology. The CPU deaths will be many I'm sure, but mine won't be one of them.

[savantu]

LOL the name is yes...the tech...well lets wait for the reviews and whether they mention where the tech actually comes from.And before you ask it is NOT AMD, you need to think where they got it from.

The only similarity between QPI and HT is that both are point to point links.Other than that everything is different , the protocol ,the transport, the link , the routing , the physical and so on.
If you want to get highly technical , I recommend you to read this : http://www.realworldtech.com/page.cf...2807020032&p=1

Work on QPI started in 2002 with Intel Tanglewood , an 8 core Itanium product.The former ex-Alpha team was in charge.
Tanglewood was canceled in late 2004 and replaced by the actual Tukwilla , yet the uncore part ( interconnect and IMCs ) was kept since it was state of the art.

By all accounts , QPI is much more sofisticated , offers far better RAS features, it was designed from the head start to be independent of the physical implementation ( it can go optical w/o problems ) than HT.
HT's main problem is that they are forced to maintain compatibility and as a result although new for the x86 world at introduction , HT and its updates are inferior ( especially in RAS) to a clean sheet design that took so long to cook up.

Can you really see Intel joining the HT consortium with AMD and NVidia??...me neither
...

Intel suffers from the NIH syndrome , that's Not Invented Here.As a result , they gladly spent hundreds of millions $ , millions of man hours and huge engineering resources to reinvent the wheel from scratch.The end results is better for their purpose , yet it took them far longer and at much higher costs than simply licensing what was available on the market.But we're talking Intel after all.
End with the off topic.

[MACMAC]

I have the boards, just need the CPU now

Glad to see that many of us are on the same boat...


It will be interesting to see how D9GTR/GTS fares at low latencies/low voltages.

[Hamilton]

We are over FSB, now dawns a new age of low bus speeds and ram tweaking a whole new way...im just waiting for all the Intel fanbois saying its like clocking ram on AMD boards...LOL

I'm completely with you on getting the most speed out of the ram but how would you overclock an Nehalem with locked multi if you don't change the FSB?

Or do we simply need to get a new word for the CPUs base frequency instead of FSB?

[Tony]

There is no FSB on Nahalem, just like there is no FSB on an AMD processor. For Nahelem you can still adjust base frequency up a little to overclock but you may only be going from 133 to 150 or so.
My point was the worthless 600FSB+ we are seeing now is gone...the tricks played to get this bus speed that 99% of end users had no idea about and the implications it had on ram vendors brought about an FSB hate within me for the past 6 months. This was beside the fact that FCG along with Raja and Gary tried their upmost to educate the masses on why 450 or lower FSB with tight tRD was faster...no one seemed to listen though.

that's validated speed that each speed bin has been tested to run at. Please don't confuse that with achievable speed by this community.

All the speed bins have memory and QPI multies unlocked

Now Blauhung is saying QPI multies are unlocked, much like they are on Phenoms HT, and memory multies (now im unsure here this could mean simple clock ratio's OR it could mean IMC clock speed...or of course it could mean both) so there are still ways you can eek out more speed here without having to push the ram to silly speeds with lots of voltage.

Im going to quote Phenom as I do not have i7 yet....there are 2 ways to push write speeds higher, 1 increase ram speed.. or leave ram speed lower and increase IMC speed. If you push ram speed higher and leave IMC speed low you are actually limiting how much performance the system will have AND limit how far the ram will OC. So if i7 is similar to K10 you will have to increase the IMC core clock to make running 2000mhz ram even worth it, BUT just increasing IMC core clock has the same effect as running 1 speed up on the memory.
Now...it may not completely work like this on i7, reviewers hopefully will explain to us all shortly, so watch out for talk of increasing IMC clock even with lower ram clocks but tight latencies maybe beating higher clock higher latency ram speeds.

Now here is the killer crunch, on desktop Nahelem may just have to much bandwidth even at 1067 or 1333, with the added width you may get the horn off the sandra bandwidth scores but you won't get the added FPS in games. So this is where tighter latencies will come into play, running 5-5-5- at 1333 over 7-7-7- may add a greater % of FPS in a game than running 8-8-8- or 9-9-9- at 1800+MHZ with higher vdimm. it may be better (if available) to increase IMC core clock sticking to 1333 ram speed and tighten latency...

All soon I hope will be revealed.

[Blauhung]

There is no FSB on Nahalem, just like there is no FSB on an AMD processor. For Nahelem you can still adjust base frequency up a little to overclock but you may only be going from 133 to 150 or so.
My point was the worthless 600FSB+ we are seeing now is gone...the tricks played to get this bus speed that 99% of end users had no idea about and the implications it had on ram vendors brought about an FSB hate within me for the past 6 months. This was beside the fact that FCG along with Raja and Gary tried their upmost to educate the masses on why 450 or lower FSB with tight tRD was faster...no one seemed to listen though.



Now Blauhung is saying QPI multies are unlocked, much like they are on Phenoms HT, and memory multies (now im unsure here this could mean simple clock ratio's OR it could mean IMC clock speed...or of course it could mean both) so there are still ways you can eek out more speed here without having to push the ram to silly speeds with lots of voltage.

Im going to quote Phenom as I do not have i7 yet....there are 2 ways to push write speeds higher, 1 increase ram speed.. or leave ram speed lower and increase IMC speed. If you push ram speed higher and leave IMC speed low you are actually limiting how much performance the system will have AND limit how far the ram will OC. So if i7 is similar to K10 you will have to increase the IMC core clock to make running 2000mhz ram even worth it, BUT just increasing IMC core clock has the same effect as running 1 speed up on the memory.
Now...it may not completely work like this on i7, reviewers hopefully will explain to us all shortly, so watch out for talk of increasing IMC clock even with lower ram clocks but tight latencies maybe beating higher clock higher latency ram speeds.

Now here is the killer crunch, on desktop Nahelem may just have to much bandwidth even at 1067 or 1333, with the added width you may get the horn off the sandra bandwidth scores but you won't get the added FPS in games. So this is where tighter latencies will come into play, running 5-5-5- at 1333 over 7-7-7- may add a greater % of FPS in a game than running 8-8-8- or 9-9-9- at 1800+MHZ with higher vdimm. it may be better (if available) to increase IMC core clock sticking to 1333 ram speed and tighten latency...

All soon I hope will be revealed.

I'm unsure about how the IMC clock and Memory clock are implimented, but the bios screen from the smackover seems to show 2 clock multipliers for memory


I would guess that UCLK would be IMC clock multi, and it's somewhat obvious what "memory multiplier" means

also I hadn't found this posted but anandtech did a nice blog about memory voltage and how they already know some tricks to optimize a board for higher voltage (just can't tell us yet, stupid NDA). Just without those optimizations is where you run into the danger zone. As with all things, if you make it possible for people to break something, they will do it. And from a business standpoint if it's easy to do and you put no warning on your stuff, people will complain.
http://www.anandtech.com/memory/showdoc.aspx?i=3426
Edit: haha, just noticed that it was posted at the top of this page

[Rammsteiner]

Now here is the killer crunch, on desktop Nahelem may just have to much bandwidth even at 1067 or 1333, with the added width you may get the horn off the sandra bandwidth scores but you won't get the added FPS in games. So this is where tighter latencies will come into play, running 5-5-5- at 1333 over 7-7-7- may add a greater % of FPS in a game than running 8-8-8- or 9-9-9- at 1800+MHZ with higher vdimm. it may be better (if available) to increase IMC core clock sticking to 1333 ram speed and tighten latency...

Also keep in mind, at least for K10, is that when you run 3:8 mode that you cant select lower latencies than 5. To make use of DDR1065+ speeds in combination with lower latencies you had to use 1:1 and thus increase the reference clock a lot.

Ive no idea if this was a design flaw or this will be on i7 as well. It's a possibility though this might be the case on i7. But since the reference clock of i7 was designed for 133Mhz, Im not so sure whether you can increase it all that much. Enthusiast boards will feature and show better options, but as you already said, maybe you'll already have to stop at 150Mhz. At least the memory multiplier shows thus far no reasons to above noted K10 situation since if 150Mhz is doable, you're still at 1500Mhz DDR3 with the same latency options. But if it stops at 150Mhz you'll have to use a higher multiplier which 'might' result in crappier latencies although the RAM could have handled it.

[zads27]

we have cas9 1800 parts 1.5ish volts...1400ish cas5/6 -5-4- is faster...so you can run at 1800 and be slow or run 1400 to 1500 with tight latency and be fast...choice is yours

There is no FSB on Nahalem, just like there is no FSB on an AMD processor. For Nahelem you can still adjust base frequency up a little to overclock but you may only be going from 133 to 150 or so.
...
Im going to quote Phenom as I do not have i7 yet....
...
Now here is the killer crunch, on desktop Nahelem may just have to much bandwidth even at 1067 or 1333, with the added width you may get the horn off the sandra bandwidth scores but you won't get the added FPS in games. So this is where tighter latencies will come into play, running 5-5-5- at 1333 over 7-7-7- may add a greater % of FPS in a game than running 8-8-8- or 9-9-9- at 1800+MHZ with higher vdimm. it may be better (if available) to increase IMC core clock sticking to 1333 ram speed and tighten latency...
All soon I hope will be revealed.

Curious Tony, what's your basis for this first statement you made, that "1400ish cas 5/6 is faster"?

Bandwidth is no problem. 1066 CL6 can net you >20GB/sec in Sandra benchmarks. As far as FPS being tied to latency.. well, sort of. It helps in some cases, not in others.

Even Gainestown ES easily goes to >150% (how much I won’t say) standard ref clock at its original multiplier.. with Nehalem, overclocking isn’t immensely different from FSB boards.. you just have more things to juggle now. You have your ref clock and then your multipliers. Not that big of a leap from FSB stuff, really. These are not my screenshots, but since its already on the internet, I’m willing to discuss it without fear of Intel taking my Nehalem processors and X58 boards away:

As you can see, there’s a 12x multiplier for memory. It’s actually quite useless if you are manually overclocking, since you really want to push the ref clock up.
The only situation where I see this doing much good is if you are running the standard ref clock and proper XMP certified memory which will make use of this multiplier.

[Tony]

Ask your self this question:

Is all the bandwidth at 1333 used by the processor? In fact is all the bandwidth at 1067 used by the processor?

Limit your answer to the use on a desktop system, not server....and be truthful.



for 99.555555555% of end users the best way and safest way for them eek out more speed is to lower the latency, if you can increase IMC clock then add this also but you can leave the ram speed and voltage alone.
There will be the 200 users here on XS who run 2000+, we all know that but 200 users here does NOT pay my salary, the money is with the other 99.55555555555% that populate the world and want to have a little extra for nothing, hence play it safe and go with subtle tweaks that gain you increases without the need for huge vdimm or vcore increases.

So yes I know running 2000 may be slightly faster, what i am talking about here is doing it in a safe 24/7 environment where people pay for the ram and the processors they use and are not sponsored or rich enough to not care. I get freebies, I think we all know that, but i try to advise caution and look after all manufacturing parties involved...I have killed more ram, more CPU's and more motherboards than most testing so i tend find out quick what is the best way to approach new tech. In this case I feel its best that end users use low vdimm, they get away from thinking they have to run 2000+ until we can produce those dimms in quantity and at a voltage that does not harm. Remember most of the guys here have Micron DDR3, this still needs high voltage even on i7 for high clock speed so advising low latency really is the way to go.

While I am waiting for the CPU's to arrive, run some tests for us showing a fixed CPU speed, differing ram speeds and the effects of latency on writes and access. Push this into some games so we can see where the FPS are going and then plot a correlation of ram-speed Vs FPS Vs ram voltage. Then we can see just how much gain there is at 2000 over say 1500 with 9-9-9- at 2000 and 5-5-5- or 6-5-5- at 1500. factor in the risk values also running 2000 with higher vdimm and then we can all look at whether its worth it. Best you use older Micron D9 for this test, not the newer (various manufacturer) low voltage DDR3 then is just being released (we are testing too ) as most jumping to i7 will do so with the ram they have to hand...not go and buy more.

Don't ask me to do this though...i will get 1 CPU and in no way am I risking killing it.

[bingo13]

Good article by Gary.

Thanks. As you know and several others here, we could write a few pages on the technical details and how to get the memory clocked up with higher voltages without affecting the IMC (to a certain point). I wanted to go further, but no need in upsetting the blue team this close to launch. I hate having a sock in my mouth for this long and seeing all the "High VDimm DDR3 will kill your Nehalem" articles.

For everyone else-

Can High VDimm damage the CPU? Yes it can, and it has, but nothing really different than what occurred on Phenom/A64 until the users and BIOS engineers were educated. Intel is playing it safe and rightfully so up front with the 1.60V~1.65V guidelines. I do not blame them as this platform will be a significant change for users weaned on FSB for the past decade.

I over simplified the HT to QPI comparisons in the blog, but really, anyone with clocking experience on the A64/Phenom platforms will feel right at home. There are a couple of additional/important settings to learn and the terminology is a little different, but I think a good A64/Phenom clocker will be about 80% there on day one.

That said, I feel really safe right now that up to 1.8 VDimm and correct BIOS settings will be fine for the general enthusiast (time will tell). For the more adventurous overclockers, the right settings will result in 2.0V~2.1V for benchmarking purposes (not so sure for long term usage as I had my hand on two CPUs today that did not survive 24/7 benching at higher than 2.0 VDimm settings). This is more than enough for current Samsung ICs, probably a bit much for the new Qimonda ICs. I am testing Micron DDR3 now and will report further on it next week.

For the rest of us, getting CAS5 at 1333 or CAS6 at 1600 in Tri-channel operation with 1.65V will result in some phenomenal bandwidth, write speeds, and latencies (see Tony's post above). So much so, it has been extremely difficult to even locate a program (commercial application) that can utilize the bandwidth Intel is about to give us. To be honest, we were at this point a few years ago with A64 in the beginning, but the jump this time is just incredible.

[JumpingJack]

Ask your self this question:

Is all the bandwidth at 1333 used by the processor? In fact is all the bandwidth at 1067 used by the processor?

Limit your answer to the use on a desktop system, not server....and be truthful.



for 99.555555555% of end users the best way and safest way for them eek out more speed is to lower the latency, if you can increase IMC clock then add this also but you can leave the ram speed and voltage alone.
There will be the 200 users here on XS who run 2000+, we all know that but 200 users here does NOT pay my salary, the money is with the other 99.55555555555% that populate the world and want to have a little extra for nothing, hence play it safe and go with subtle tweaks that gain you increases without the need for huge vdimm or vcore increases.

So yes I know running 2000 may be slightly faster, what i am talking about here is doing it in a safe 24/7 environment where people pay for the ram and the processors they use and are not sponsored or rich enough to not care. I get freebies, I think we all know that, but i try to advise caution and look after all manufacturing parties involved...I have killed more ram, more CPU's and more motherboards than most testing so i tend find out quick what is the best way to approach new tech. In this case I feel its best that end users use low vdimm, they get away from thinking they have to run 2000+ until we can produce those dimms in quantity and at a voltage that does not harm. Remember most of the guys here have Micron DDR3, this still needs high voltage even on i7 for high clock speed so advising low latency really is the way to go.

While I am waiting for the CPU's to arrive, run some tests for us showing a fixed CPU speed, differing ram speeds and the effects of latency on writes and access. Push this into some games so we can see where the FPS are going and then plot a correlation of ram-speed Vs FPS Vs ram voltage. Then we can see just how much gain there is at 2000 over say 1500 with 9-9-9- at 2000 and 5-5-5- or 6-5-5- at 1500. factor in the risk values also running 2000 with higher vdimm and then we can all look at whether its worth it. Best you use older Micron D9 for this test, not the newer (various manufacturer) low voltage DDR3 then is just being released (we are testing too ) as most jumping to i7 will do so with the ram they have to hand...not go and buy more.

Don't ask me to do this though...i will get 1 CPU and in no way am I risking killing it.

Tony -- thanks, some level-headed commentary from a memory guru...

jack

[Charles Wirth]

Ive stayed out of it, people will find out what is comfortable to their cooling so being sub zero has its benefits.

[NH|Delph1]

You will want to run your modules at high frequencies, but even more so, overclocking the CPU gives enormous boosts in bandwidth.

Don't worry about the memory overclocking, read the manual and you will be set to run your modules at comfortable speeds.

//Andreas

[Nikolasz]

Most? Im sure most of it is 1.5V

mine is 1.5v, but i put 2000rpm fan on my stickers, got no need for ocing