Original Link: http://www.anandtech.com/show/2712
Overclocking Core i7: A Tale of Two Retail Processors…by Rajinder Gill on January 22, 2009 12:00 PM EST
- Posted in
To facilitate our Core i7/X58 motherboard testing, we have been snapping up retail CPU’s from a variety of outlets in the US and Europe. Since most of the users adopting this platform are opting for a Core i7 920 as their mainstay processor, we have been on a buying spree for this particular model while also dropping some hard earned cash for a couple of the 940 and 965XE retail processors.
I purchased my first 920 over three weeks ago from SCAN UK. This particular CPU is a 3838A batch processor, which I proceeded to test in the Foxconn Bloodrage and EVGA X58 SLI motherboards. It’s the same old story really when it comes to clocking, I scoured the forums, spoke with Gary and Kris (who had a head start on me with this platform) and asked them the lowdown of what to expect in terms of voltage related scaling hoping that I had something at least as good if not better than their retail processors.
Like many others out there, I was hoping for a processor that could clock to 4GHz and beyond at fairly low processor core and VTT voltages. This together with an IMC that could handle triple channel speeds well above 1600MHz, actually, make that 2000MHz. The kind of speeds I’d seen posted on various overclocking forums using the 920 processors had voltages just tolerable enough for 24/7 use. I must also add at this point that some of the testing we do on these boards falls well outside the 24/7 bracket. An ever increasing number of boards are released solely with benchmarking in mind, a growing segment of the industry that seems to be on the fringe of demanding complete segregation from mainstream products due to its specialized needs.
A large part of testing these boards involves running both entry-level and high-end components well out of specification ranges, just to see what a particular board can do if pushed or if it will go up in a ball of smoke, something Gary seems to master at times. This may seem trivial to most of our readership, but it’s an ever growing part of the enthusiast sector and somebody around here has to test it, so in comes my need for decent components, especially processors in this case.
While this processor matched the ‘international average’ for core voltage scaling, I discovered that I had to apply ridiculous levels of VTT/Uncore voltage to get it to boot at anything over 185BCLK. A 3.8GHz clock speed fell with relative ease, but pushing up to 4GHz, well that’s where things became real tricky. This processor seems to stop in its tracks with VTT voltage levels over 1.36V (just outside the warranty maximum by the way) or so, causing both of the motherboards to halt during POST with a C1 error.
Manipulating voltages within the OS using motherboard specific software tools can circumvent this condition to an extent, but the voltage has to be ramped up in small steps. Either way, 4GHz stable on this particular processor is way more hassle than it could ever be worth in a 24/7 system so I am stuck at 3.8GHz. Ok, this seems a bit demanding of me, a free 1.2GHz overclock from stock and I’m nowhere near happy! In fact, at this point I was pretty much convinced that both of the motherboards I was testing were duff, not the CPU, especially when I looked over at some of Gary’s early results on the same boards, as well as results of forum members.
I conferred back with Gary on his results, he’s got three retail 920s in his repertoire (with a fourth on its way) and his results are erratic in this department too. One of his processors needs high levels of Vcore to make 4GHz possible, well in excess of 1.50V, and refuses to POST with Bclk ratios set higher than 200, regardless of VTT/Core voltages. The other two (3837A)are better than mine for IMC VTT scaling and can also handle 6GB triple channel memory at 2000+MHz with a little persuasion. Something I found impossible on my 3838A processor with 3GB of memory let alone 6GB. Also, both of his processors allow for 3.8GHz operation at stock or below stock core voltages with VTT near 1.15V. Both processors allow clock speeds to reach about 4.3GHz on 1.45V of VCore, but VTT required is near 1.425V on air cooling with a 2:8 memory ratio, change the memory ratio to 2:10 for DDR3-2000 and VTT requirements hit 1.50V, which also happens to be the maximum amount his processors will allow before throwing up a C1 code on POST.
Based on this, I decided to purchase another 920 from a different retailer in the UK, again purchased as a consumer – so no cherry picked possibilities. This one was an OEM unit, 3835A batch, a little earlier in the production timeline. Onwards to testing using the exact same components as before and much to my surprise, 4GHz + on this particular CPU is not an issue. Further, the VTT required in comparison to the 3838A processor is miniscule.
Take a look at the voltage screenshots below – VTT required for an effective BCLK of 210 Prime95 8 thread stable is a mere 1.20V under full load, Vcore in the region of 1.332V (real) for just shy of 4GHz CPU speed. If I stick the 3838 retail processor into the same board on these settings – it won’t even post, in fact I can barely get it to post at 195BCLK using 1.36VTT, which then proceeds to throw a BSOD as soon as the system is presented with a load – regardless of how much I tinker with voltages.
Bottom line here is that the integrated memory controller in the i7 Core processors is highly variable. We’ve seen results much better than ours in some instances, and we've also heard of users finding it hard to get processors to post much beyond 3.6GHz. Based on our research to date, lot codes 3835, 3836, and 3841have the best opportunity of receiving a processor that will overclock well while offering both excellent VCore and VTT voltage rates. It is still luck of the draw for the most part, but chances are these lot codes will offer improved clocking rates at lower voltages.
This type of variation in the IMC will probably make some of the upcoming CAS 7 rated 6GB 2000MHz DDR3 kits a bit of a minefield for tech support staff. Personally, we’ve not managed to crack 2000MHz unconditionally stable on our 920 processors yet at CAS7 when increasing the BCLK to 200 and running a 2:10 memory ratio. CAS 8 works, yes Gary’s got from a-b in that regard, but CAS 7 seems like it’s a no-no thus far unless you’ve got a really voltage friendly IMC or get stupid with VDimm. It’d be fair to say this is where the higher multiplier processors on the 965XE are a definite advantage.
Having a strong IMC and motherboard combination that can run the 2:12 memory ratio with Bclk at 166 is ideal for 965XE users needing both high core speeds via the CPU multiplier and memory at or above 2000 for benching. For the rest of us, we think running DDR3-1600 at CAS6 with the upcoming kits will offer the best blend of performance, stability, and cost. Just remember, BCLK is nothing but a reference clock on this platform, the only important element to pay attention to is that QPI frequency is derived from this reference clock. When you’re pushing a lower multiplier processor at higher BCLK levels you’re fighting against IMC QPI frequency limits and the ability of the IMC to swing voltage into a capacitive load at CAS 7 with DDR3-2000 speeds; tricky to say the least. This is also one of the reasons why lowering your memory ratios results in either higher processor clocks and/or additional stability at lower timings. It is simply the system reducing load on the IMC and is another reason why getting 12GB stable at DDR3-1600 is so difficult without a decent CPU and top quality motherboard/BIOS combination as one example.
So let’s say you have a really good processor, motherboard, and memory by this point, you're home free right? Sort of, there is one last caveat we found while overclocking that is worthy of mention; the accuracy of reported voltages in comparison to what the processor really receives from the board, especially VTT/Uncore voltage. We’ve noticed a number of our test boards that read BIOS or OS voltage readings close to the output side of the VTT PWM inductors. As VTT is specified to draw high levels of current, readings taken at various points across the power plane board vary, sometimes greatly; especially if the amount of copper used for traces is thin under the CPU socket.
Intel specifies a combined supply for VTTA and VTTD (the analog and digital voltage portions of the IMC) for the Nehalem architecture. Controller level droop is also part of the specification due to power consumption. Most vendors are suitably over engineering the VTT rail on their overclocking boards by using a 2-phase PWM circuit which gives a more than acceptable 0.02V~0.05V droop depending upon load. This is nothing new or evil for those that understand the nature of transient recovery, overshoot and ringback when it comes to high current power supplies.
Because the all important VTTA supply is coupled to VTTD (for various reasons), any voltage droop during power delivery affects the output voltage swing from the IMC transmitter, which is one of the reasons we see exotic controllers being employed on most of the enthusiast level boards to facilitate getting that last few percent of overclocking stability. Low noise, low impedance power is pivotal in getting the best overclocking potential from the IMC.
Although these power delivery precautions are being taken on the controller and decoupling side of things, we’ve noticed that some boards appear to be skimping when it comes to copper content on the under socket VTT power plane. Which does raise a few question marks unto the validity of going to the trouble of using exotic PWM controllers. Trace losses on some boards are in the region of 0.15V from the last VTT decoupling cap to the under socket grid when the CPU is overclocked and under full load (especially the north side of the socket). This may not seem like a big deal at first, but many users have been duped into thinking their processors are not clocking well when setting VTT to recommended values, without realizing that software related voltage reporting is way off the mark compared to the actual voltage the processor is receiving.
Many vendors are now including measuring pads on their boards that can easily be used with a DMM for voltage monitoring – a good thing, providing those pads lead back to an area of the power plane that best represents the voltage that a component actually sees at its rails. We think this is one area of the first generation of boards that needs to be addressed; accurate voltage monitoring and perhaps a little more importance given to ensuring that VTT trace losses are kept to a minimum by using a suitable amount of copper in the VTT power plane.
We modified an X58 board here in the labs with a section of thick 18 gauge wire to see how much of a difference it makes in terms of overclocking when trace losses are reduced. For the most part, the only benefit in doing this turned out to be efficiency due to better power delivery rather than outright overclocking gains. However, our ability to understand the actual VTT voltage being delivered did allow us to compensate for any voltage differences in the BIOS. This ultimately led to higher overclocks as we were able to properly setup the board.
In the end, overclocking success with the Core i7 ultimately is dependent upon the quality of the IMC much more than the motherboard. Certainly more so than we originally believed and of much greater importance than it was during our years with the Core 2 series. We will be back in the near future with additional details on overclocking the i7 series and updated results on each X58 motherboard including actual versus reported voltages.