Nehalem and memory overclocking

[Andreas G]

We've posted information regarding Intel's Nehalem platform and its support for various DDR3 memory clock frequencies in the past. We've come across some more now, actually a lot more. As stated in the previous reports, the Nehalem platform will officially support only DDR3-800 and DDR3-1066, but there is still a strong buzz that there will be official support for DDR3-1333 memory, but it has not been decided yet.

There have been some confusion as to why Intel Core i7 would not support DDR3 memory clocked at higher frequencies and the first thing you need to be aware of is that it is not the frequency of the memory that is the culprit.

The problem is rather that the memory and the processor internals are fed synchronized voltages. Previous reports have stated that anything above 1.65V would fry a Nehalem processor and it would certainly be hard to get any of today's DDR3 memory modules to any kind of decent speeds at this voltage. But, these stories are exaggerated, but the fact remains that Nehalem processors are not going to like voltages above 1.7V.

You have to put it through some extreme handling for it to give up right away, but if you feed the memory with voltages above 1.8V, the processor could start degrading and finally cease to work. According to the information we've received you need to hit close to 2.0V to kill a Nehalem processor.

The conclusion is that the Nehalem platform, with its synchronized CPU and memory voltages, will be limited by the DDR3 modules ability to reach high frequencies at lower voltages.

This should create a whole new niche for DDR3 memory since we've also received information that says that latencies play only a minor role with Nehalem. More aggressive timings will boost scores in pure memory benchmarks, but the generic system performance seems to care little, if anything at all, about timings. This could mean that memory manufacturers could start to focus on hitting high frequencies at low voltages and loose timings for our dear Nehalem processors.

Already, tests with Corsair's DDR3-1333 memory that handles 1600MHz at 1.64V and 1800MHz at 1.7V have been performed so we should still see some decent memory frequencies on the Nehalem platform, but it will certainly be a more delicate process of reaching there than it has been before.

[KTE]

OK, interesting Andreas. This sounds like single plane voltage control between the CPU/IMC...

So if I decreased the vcore to 1.3v, does that mean the IMC/RAM is also being fed 1.3v and no high or lower?

Even if they were linked, do they not provide ratios to choose voltages, i.e.

CPU Voltage - 1.10v
IMC Voltage (CPU:IMC) - 1:1, 1:1.25, 1:1.50, 1:1.75, 1:1.20 ... upto 1:1.8
DDR Voltage - same as above

?

If they don't provide ratios it means they are using the same voltage chip for both voltages.

Also, onto what you mention next - if Nehalem is similar to K10 IMC wise, IMC/DDR frequency is all that will matter, not latencies. Meaning 100MHz higher on these will outrun 100Mhz lower with very tight latencies.

[Delph1]

I don't know if it's that easy, and I was cautious to write CPU internals and not core voltage. I believe the core voltage and memory voltage are separate.

No ratios as far as I know.

Yepp, latencies are fading in importance

//Andreas
_________________
"I'm sort of a postmodern vegetarian, I eat meat ironically"
"You are our Swedish superhero "

Forget everything, knowledge is your enemy, it's only going to make you second-guess yourself."

[KTE]

Ah, true. Internals are characterized by core/uncore now so it can be tricky. Good point

AFAIK, the CPU is fed synchronous voltage to the memory and >1.65v is the limit till you kill the CPU using air or water but not more cold. I have no evidence if that means the actual core or just one of the dozen parts of it, but seeing the power technologies involved with Nehalem, I am pretty much certain in practical working the CPU core voltage cannot be the same as the memory voltage since that would really fuzz up the whole power saving technologies and overclocking, top to bottom. It just won't work this way, so its nearly impossible to have been implemented IMHO. I think the confusion stems, as did with K10, that the synchronous voltage is probably IMC-RAM or one of their components related with a tricky IMC voltage limit and a dead IMC means a dead CPU since it is now inbuilt. I don't think 1.65v on cores will kill it but that amount on an IMC killing it is not out of question. Anyway, a month or so and time shall tell

I'm still hearing of Core i7 900 EE, Core i7 900 and Core i7 800 shipping for some reason since a few months now.

[Mean Machine]

I guess the overclockers will soon have a reason to return to AMD and AM3. They way I see it, a lot of people will fry a lot of CPU's without getting near the max potential for the RAM, which is a shame.
_________________
Man, or woman, or both?
Grandis spiritus, sanctus diavolos

[KTE]

I doubt that MM.

Not if 45nm C2 is still around to clock to 6.6G and neither if the new 45nm clocks to >4G with superscalar 8 threaded core. The numbers oc'ers want, 3DM/Pi/CB10, will all be dominated by the new Intel 45nm, so I see them sticking to it safely. Cost etc has never been something they think of anyway

I think Intel knows very well that the main market demand and domination even in oc will still be C2D 45nm and thats why most of Nehalem is well delayed to 2H '09. I also estimate they will speed up the intro of Nehalem based parts if the new AMD 45nm isn't as good as they expect - note: as threatening to their 45nm as they expect from hype (15-30% over 65nm in server field). Thats typically how they work.

[Mean Machine]

I'm aware of that, but I see the memory bandwidth going the wrong way compared to AMD at least. I just hope they can pick up the pace with AM3.
_________________
Man, or woman, or both?
Grandis spiritus, sanctus diavolos