Core i7 overclocking: Uncore frequency explored

[Aivas47a]

One of the features of the new core i7 architecture is uncore frequency.

Uncore sets the frequency of the on-die memory controller and the L3 cache. Like cpu clock speed, dram speed, and QPI frequency, uncore is a multiple of Bclk. Uncore can be set independently of those other frequencies, subject to certain stability limitations.

To isolate the effect of uncore, I ran some tests holding BCLK steady but gradually increasing the uncore multiplier. (For these tests, BCLK is 211 and the DRAM multiplier is 6x.)

First, comparing Super Pi 32M times at different uncore multipliers (ratio to DRAM multi) yields the following:

12x uncore (2:1) -- 10m 6.953s
14x uncore (7:3) -- 10m 00.29s
16x uncore (8:3) -- 9m 52.01s
18x uncore (3:1) -- crash
20x uncore (10:3) -- no boot

(Note: raising the DRAM multi from 6x to 8x stabilized the 18x uncore, yielding a 32M time of 9m 36.481s).

Similar improvements in memory and L3 cache performance can also be seen with the Everest cache and memory benchmark.

12x uncore (2:1):



14 x uncore (7:3):



16x uncore (8:3):



18x uncore (3:1):




Notice that increasing uncore yields significant improvements in memory read, write, copy, and latency performance. L3 write and copy times also improve but -- interestingly -- L3 read does not.

Uncore overclocking tips:

To summarize uncore tweaking:

1. Uncore must be in at least a 2:1 ratio with DRAM frequency. Thus, using a 6x DRAM multiplier above, 12x is the minimum uncore multiplier.

2. Increasing the uncore:dram ratio above 2:1 yields significant performance gains. However, when the ratio reached 3:1 it was not possible to maintain full stability.

3. The uncore:dram ratio is more determinative of stability than the raw uncore speed. For example, the 18x uncore multiplier was unstable when DRAM was set at 6x (3:1), but raising the DRAM multiplier to 8x (9:4) restored stability.

4. To stabilize uncore at ratios higher than 2:1, try increasing the VTT PWM Switching Frequency. (Note, not all bios versions expose this setting, and it may be presented differently by different mobo manufacturers.) For example, stable operation at 7:3 required a 15% increase in VTT PWM, while stable operation at 8:3 required a 30% increase.

In sum, uncore tweaking may not be too important for 24/7 operation, but for benching (or just to get the maximum performance from your core i7 machine), uncore is definitely worth exploring. (Note, one of the touted advantages of the upcoming EVGA Classified X58 board is its robust uncore capability. We'll keep an eye on whether that pans out.)

Thanks for reading.

[brentpresley]

GREAT findings!


//goes to re-read and memorize//

[UncleWilley]

Looks like a lot of work. Great job!

Now I just need to figure it out.

[CHSIsupplier]

Awesome job Aivas! Very interesting results.

[Aivas47a]

One other tidbit I meant to mention: I've seen the comment made in various internet forums that uncore frequency can not exceed QPI frequency. However, in my testing I did not find that to be a limitation.

[UncleWilley]

Came across this on another forum. Didn't know if this is anything new.

Shocker! Retail Core i7 CPUs caught using DDR3-1600 memory

The kids over at Maximum PC have told us a sordid tale involving Core i7 processors, "extreme heat and smoke damage," and a pair of jumper cables -- we'll spare you the details -- and have come to the realization that, contrary to popular belief, the retail version of the budget-minded Core i7 920 processor (and its midrange brethren, the 940) is shipping with unlocked multipliers and memory ratios. According to some "Deep Throat"-esque shadowy figure at Intel, the company chalks it up to a "marketing decision" made after receiving "requests from some of our customers." If this is indeed the case, customers will be able to rock the fast-paced DDR3-1600 memory without shelling out big money on a Core i7 Extreme Edition 965 CPU. And that ain't the least of it -- hit the read link for plenty of talk about QPI speeds and benchmarks. You'll be glad you did.

Exclusive: Retail Core i7 CPUs More Powerful than Originally Reported

[Aivas47a]

That's kind of an odd story ... the mag makes it sound like they discovered a big secret, but everything they're saying has been known since core i7 was released.

[UncleWilley]

Originally Posted by Aivas47a

That's kind of an odd story ... the mag makes it sound like they discovered a big secret, but everything they're saying has been known since core i7 was released.

I guess it's all about headlines just like newspapers and TV news. They get attention with a headline like "Shocker". Even if it's old news they bet at some people will "stick around" and read more of their paper etc. Also, why bother to check out a stories accuracy if it will "sell papers".