Intel’s Core i5-655K & Core i7-875K: Overclocked and Analyzed
by Rajinder Gill on May 28, 2010 5:00 AM ESTBetter than our i3-540 for Overclocking?
Part of the attraction generated by the allure of unlocked multiplier processors is the additional flexibility purportedly provided, allowing any manner of clock domain frequency combinations. The obvious question has to be what an unlocked multiplier set offers the Clarkdale topology? The real ticket to retaining performance on this platform is to keep QPI frequency as close as possible to processor core frequency. That’s because the memory controller is off-die and the CPU communicates with the memory controller via the QPI link (the data link between the MCP stages in the images below).
Too wide a gap between both clock domains and the CPU must hold data for longer periods of time in the associated buffer stage before it can be transferred to the next bus. As a result, setting too high a CPU multiplier with too low a reference clock (BCLK) would neglect QPI bus frequency. This is because Intel limits the range of QPI multipliers at our disposal. What we want to use is a CPU core multiplier that is the same as or one notch higher than the QPI multiplier ratio.
Another thing that becomes readily apparent with the Clarkdale topology is that memory speeds around the DDR-3 1800 mark are generally the upper limit for most processors when absolute stability is required at reasonable voltage levels. Slightly higher speeds are possible but only by sacrificing valuable memory access latency – meaning you have to either run loose memory timings or drop QPI frequency to compensate. Given the choice, we’d sacrifice a little memory bandwidth before sacrificing a lot (lowering the QPI frequency). With this in mind, all of our comparisons have been made using the highest available QPI multiplier ratio on each processor, with the actual QPI frequency closely matched:
Facebook
Twitter
RSS
51 Comments
View All Comments
Rajinder Gill - Tuesday, June 1, 2010 - link
As the graphs state - this is VCC/VTT power only, the two major power rails of this architecture. The 12V ATX fan headers and PCIe 12V only on the E659 motherboard. Power to DDR3 is not something I focused on but may do in a future piece (there will be a frequency proportional rise in power provided timings are not changed). A very crude guess - I'd expect the rise over stock to be around 5 watts on the DRAM side in this frequency band (and total draw to be no more than 10~15w). Although figures would differ according to the scaling capabilities of various modules.Other than that, there's not much else aside from CPU PLL which is specified at around 1.1 amps at 1.8V (around 2-3 watts tops).
Rajinder Gill - Tuesday, June 1, 2010 - link
EDIT: The 12V ATX fan headers and PCIe 12V only on the E659 motherboard.That should read the 12V ATX line supplies fan headers and PCIe 12V power only. The 3.3v and 5V rails supply DDR, CPU PLL, IOH (and all derivatives such as IOH PLL, SB~IOH termination voltage etc).
Later
Raja
DanNeely - Tuesday, June 1, 2010 - link
Time permitting I'd be interested in seeing those numbers as well. I understand your desire to measure power consumption closer to the source. My concern is that increased power consumption from the secondary items you're not measuring is a black box; while the AC-DC conversion loss in the PSU from measuring power at the wall can be mostly corrected away by looking at what the efficiency rating of the PSU used in the test setup is.DanNeely - Tuesday, June 1, 2010 - link
Time permitting I'd be interested in seeing those numbers as well. I understand your desire to measure power consumption closer to the source. My concern is that increased power consumption from the secondary items you're not measuring is a black box; while the AC-DC conversion loss in the PSU from measuring power at the wall can be mostly corrected away by looking at what the efficiency rating of the PSU used in the test setup is.Rajinder Gill - Tuesday, June 1, 2010 - link
I prefer to keep things at the DC level. There are plenty of articles covering wall level consumption with your standard kill-a-watt type unit (and they also state the PSU used so users can factor out the losses if they know the effective efficiency curve). I think you are worrying too much about the lesser rails. Sure they will make interesting reading at some point - but there is nothing that pulls more than a couple of amps so the effects on power consumption will not be huge.Regards
Raja
Rajinder Gill - Tuesday, June 1, 2010 - link
Ok "DanNeely", this is for you,I just ran tests on the 3.3V and 5V rails. At stock the combined power consumption of these two rails on the E659. Bear in mind this is an enthusiast level board (higher switching losses due to higher switching speeds on VDIMM, plus using an NF200 for PEG multiplexing):
Running DDR3-1333 CAS 8-8-8-24 with 4GB of memory. (3.3v + 5V rails combined).
20.2 watts idle
Linpack load = 26.36 Watts.
That's a change of 6 Watts between idle and load.
At 4.551GHz, now running DDR3-1820 (1.60 VDIMM):
Idle = 22.86 Watts (2 Watt idle increase)
Load = 27.86 Watts.
That's a 1 Watt increase over stock speeds under load with an overclock of 1GHz on the CPU (running QPI over 4GHz). Hardly worth writing about. Do note - the effective change will vary from board to board according to VRM switching efficiency (which is coming into play if you look at the deltas between idle and load). Of course, I am not including things like HDD's etc although, some of the static 3.3V and 5V rail consumption is due to the GPU (GTX 275) which also draws a little power from that rail.
Hope that answers your questions.
Regards
Raja
DanNeely - Tuesday, June 1, 2010 - link
yes it did. thank you.tno - Friday, May 28, 2010 - link
Setting aside typos, I know wall of text is to be avoided but this felt almost like the opposite problem. Additional clicks for additional adviews. Clearly you didn't like what you saw out of the chips despite the voltage improvements evident in the more modest chip. Yet in the end despite devoting 1/9th of the coverage to it, you reward your recommendation to the pricier chip just one page after showing it severely underperforming its non-K analogue in both overclocking and voltage.I have said it before and I think its worth mentioning again, clearly there is a lo of passion for tech in the growing AnandTech team, but maybe adding a team member whose passion is writing and across whose hands every article will pass would give the site that extra polish that elevates it from other tech sites.
Jason
7Enigma - Friday, May 28, 2010 - link
At first I thought the same thing (saying nothing at all in the final page about the 655K does leave me puzzled as it is good), but Rijinder did clarify his recommendation for the 875K by saying PRICE. He is saying for the price of these chips and where they fall in line with the rest of the offerings from AMD and INTEL, the 875K is in a sweet spot. Remember the 655K is a dual-core,4 thread chip for $215 (lots of competition from both camps), while the 875K is a quad-core, 8 thread chip for under $350.In the end (and after re-reading the conclusion and article) I think the last page needs to have a bit more meat behind it. The data in the article itself is very detailed, but the final wrap-up needs some work. But honestly, since it's a free site, I'll take the good data and sketchy conclusions (I tend to make my own).
Thanks for the article!
troun - Friday, May 28, 2010 - link
"Past 3.9GHz, we’re already looking at a 10W increase in power consumption for every 20 MHz rise in CPU frequency"
But I read 20W for 100Mhz (or 10W for 50Mhz), with ~160W @ 3.9Ghz and ~180W @ 4Ghz...
However very interesting article, a similar curve (W/Mhz) would be also appreciate for an i7 9x0 comparison, 32nm Vs 45nm (980 Vs 930?).