Core 2 Endurance Battle pits Nvidia 680i versus Intel 975X

SINCE ITS ARRIVAL last month, right on my birthday - don't ask for my age - the Nforce 680i has seemingly established itself as the absolute top end of Intel PC chipset platforms, at least until AMD's last Intel chipset effort, the RD600, shows up in limited DFI and Asus entries after the holidays. Its features, performance and, especially, overclockability are supposed to clearly lead ahead of Intel's current offerings.

Most top mobo vendors from Taiwan have both of these platforms in their sales brochures, with some of them also adding high-end offerings based on Intel's P965 (it seems to have a bit better results with Core 2 chips) and of course the upcoming RD600. So, most of them will avoid direct confrontational style comparisons - relationship reasons come to the mind too, of course. If the vendors won't do it, well - we will.

At the same time, the quad-core 'Kentsfield' Core 2 Extreme QX6700 is officially, by far, the fastest single-socket PC CPU for 2006 (AMD 4x4 doesn't count as it is basically two modified Opterons sold for the price of one Athlon FX - it's still a dual-socket machine), but most of the real sales are the dual-core Conroes in various incarnations. I happen to have two of these, a X6800 and E6700, which came in the same batch originally, and was tempted to test them simultaneously on identical platforms, and see which one goes higher. Well, what a better opportunity then getting them to face off on two great high end boards - MSI's 975X Platinum PowerUp, one of top 975X boards (it's previous version was the one where nitro-cooled Conroe reached 4.8 GHz and 1.7 GHz FSB even before launch!), and ECS PN2-SLI2+ Extreme, the high-end version of Nvidia 680i reference design. Importantly, both of these boards have comparatively good North Bridge cooling - MSI has a large all-copper heat sink to which I added a small (but high-RPM) fan, while the ECS entry has Nvidia's standard large heat sink and unusually mounted matching side fan.

On top of it, the proliferation of ridiculously fast memory modules continued - my latest regular monthly Taipei round brought along the freshest in Taiwan's 'enthusiast' module entries, the 2x1 GB GEIL DDR2-1160 CL 4-4-4-12 and G.Skill DDR2-1066 CL 4-4-4-5 modules, good to square off against my old trusty reference, the (also 2x1 GB) Corsair XMS6400CL3. Again, let's run them all on both chipset platforms and see how high they fly. No, it's not just about the absolute megahertz, but, more so, about getting the minimum latency at minimum voltage (read: the least heat and power usage) when the memory runs truly in sync with the FSB. Often, this produces equal (or better) results than trying to French fry the modules at ridiculous frequencies (and matching voltages) which result in bandwidth well beyond what CPU FSB can take, but at the cost of latency.

At the end, I pulled out from nowhere one of Zalman's hard-to-get and very expensive 9700CNPS all-copper blue-LED air coolers with a 'nail polish brush' thermal grease bottle (allows very thin and even grease application on the CPU heat spreader), added the ECS GeForce 8800GTX graphics, and connected everything to the old MGE 500W power supply. All the stuff ran Micro$oft WinXP SP2 with the usual stuff on it - no, I am not rushing into Vista as I don't like DRM infections and yet another round of security hole adventures; more likely to adopt some form of Linux benchmarking add-on in the future. By the way, the test was performed in my hot Singapore test lab at around 32 C room temperature with all air conditioners and room fans turned off.

The test was also done in non airconditioned test lab at 32 C room temperature without any room fans.

So, here we got a potentially interesting competition: CPUs (X6800 vs E6700), chipsets (680i vs 975X), boards (ECS vs MSI) and memories (GEIL vs G.skill vs Corsair). Only the cooler, GPU and PSU are shared (OK, so are the 20-inch Philips UXGA LCD, HDD and DVD-ROM, too). Let the games begin!

Board comparison
Being both high-end entries, these two mainboards come with full fledge features and associated bells & whistles: fashionable black PCBs, all the possible interfaces on board, including a zillion SATA and USB ports (how come there is no more FireWire 800, the 1394B interface, on board anymore?), enormous power regulation blocks with extra passive cooling, and well cooled North Bridge and South Bridge devices. Complete sets of cables, thick manuals and CD sets complete the colourful packages.

That's where the similarities end: let's start from the BIOSes, which couldn't be more different. The Nvidia BIOS on the ECS board, done by Phoenix - Award, has probably the most extensive setting, fine tuning and overclocking options available, down to every single voltage, DRAM timing parameters, even saving up to three overclocking profiles for different occassions. Not to mention the status display and dedicated on-board power & reset buttons - this board's features were reviewed in more details last month in our Kentsfield Nforce review.

The MSI BIOS is of the same Phoenix - Award progeny, but the array of features is somewhat disappointing for such a high end board. In fact, in many ways it is below Intel's own D975XBX2 when it comes to the overclocking BIOS settings. It's kind of pity, since MSI uses CoreCell chip for dedicated, hardware assisted overclocking and system load monitoring, and far more could be provided to fine tune such a feature. For instance, there are no DRAM timing settings beyond the four usual latency parameters, and FSB or North Bridge voltage can't be set manually.

In other features, Nforce 680i provides hardware-accelerated dual Gigabit Ethernet with firewall functionality and, of course, nearly double the number of PCI-E lanes from both north and south bridge devices, connected via HyperTransport.

CPU & chipset shoot-out
The rule was simple: both CPUs on both chipsets, with the same spread of voltages, frequencies, FSB clocks etc applied - to ensure apples-to-apples (not Apples) comparison. As Core 2 Duo E6700 is supposed to be a lower-end chip, I ran it first on each board. On the MSI mobo, there was no ability to set the clock multiplier lower than the default 10x with this chip, so I could only adjust the FSB to get the clock up. I went ahead, and managed to get to 3.6 GHz with FSB 1440, all the way running Windows benchmarks like Sandra fine and without a hitch. When the 3DMark06 CPU test time came, the benchmark gracefully crashed (luckily, back into Windows, not the boot). The additional voltage up to near 1.34 volts didn't help on this, neither did DRAM timing relaxation and voltage adjustment. When I dropped the clock to 3.5 GHz sharp at FSB 1400, all went well at 1.32 volts, including full 3DMark06 suite completion at various resolutions. This is a good 30% proportional CPU and FSB overclock for true 30% extra performance with near default voltage, at simple air cooling and hot summer room environment, all from a pretty affordable CPU!

Again, 3DMark06 CPU is, to me, one of the most suitable benchmarks around to fully stress the CPU and mainboard, as it will drive fully all the cores available, both integer and FP portions, thrash the caches with its huge code and data sets across multiple threads, and therefore load the FSB and memory very well to use whatever you throw at it. So, my rule is that, if the configuration runs this without a hitch during the day in a hot stuffy room - it is good enough to go 'production'.

Now I dropped in the faster Core 2 Extreme X6800, the same one I managed to run only at 3.2 GHz on Intel's D975XBX2. Well, no better news there - the MSI board booted Windows with it at 3.33 GHz CPU / FSB 1333, but Windows crashed back to the boot, the moment I ran anything, including Sandra CPU test. The 10x multiplier 3.2 GHz CPU / FSB 1280 ran fine in Sandra at default CPU voltage, but 3DMark06 CPU crashed back to Windows this time. The only fully stable configuration was, again, 3.2 GHz CPU and FSB 1066 using 12x multiplier - uughh, only a 9% overclock! I really hope that my Conroe XE is an aberration, otherwise if would be quite a shame if the 40% cheaper standard Conroe variety trounces it totally in the overclocking department.

On the ECS Nvidia board, I expected to get better FSB settings, even if the CPU overclocking is limited. Unfortunately, despite all tries and adjustments, manual or auto, there was no way to make the Conroe E6700 boot as 3.6 GHz CPU / FSB 1800, or 3.5 GHz CPU, FSB 1750. Oh yes, from this you can notice that I am ABLE to change the standard Conroe multiplier downwards (but, of course, not upwards) on this board - a truly huge benefit for this class of mobos. At the end, funnily, the thing ended up working at exactly the same speeds as on the 975X, 3.5 GHz CPU / FSB 1400, for stable 3DMark06 CPU run - not a bit more (well OK, I didn't really go in tiny steps, MHz by MHz, OK?). The CPU voltages required for stable operation were still at the same around 1.31 - 1.32 volts.

As previously mentioned, I believe I could have had faster FSB and North Bridge operation here if there was more powerful cooling system for this boiling-hot circuitry. Also, Nvidia could pick up a bit of tricks from Intel's "Enhanced Power Slope" on its 975XBX board series, which somehow enables those CPUs to run at the same high overclocked speeds yet uses voltages LOWER than default (remember my Presler and Conroe reviews on these boards?). However, as you'll see from the benchmarks, Nvidia did have an advantage at the end.

Warm memories
Yes, all memories were warm, but not hot, to touch, at the default DDR2-700 in sync 2:1 clock when used with the 3.5 GHz Conroe with its x10 multiplier - same at DDR2-720 in the 3.6 GHz run. At 1050 and 1080 MHz high throughput setting, the story was different: to get the 4-4-4-8 latency from both GEIL and G.Skill, I needed to up the voltage to 2.2 volts on 975X, and 2.3 volts on Nvidia - in either case, there was no need to go to 2.4 ot 2.45 volts as stated on the vendor's stickers. The Corsair memory managed to get still respectable 4-5-4-9 on the same 2.3 volts at 1050 MHz, but for some reason the MSI system refused to boot with it at all on 1080 MHz setting - I could only use the other two memories for that run. The Nvidia system ran it fine, though.

Neither of the DIMMs could do better than CL 3-3-3-5 (Corsair: 3-3-3-6, the last number isn't really that critical, but still...) at DDR2-700 speed all the way up to 2.1 volts. I'd have loved to have CL 3-2-2-5 there instead, but, well, the above latency at generic 1.9 volts and stable operation across the most demanding tests is, to me, a good use of the 11.2 GByte/s peak bandwidth provided by the FSB 1400.

All three DIMM kits use visually similar standard heat spreaders - no fancy ultralarge finned stuff here, not to mention water cooling. However, the G.skill model has the most open top side for potential extra airflow under the heat spreaded. When I mounted the Evercool memory DIMM cooler, a funky two-fan crab-like appendage, on the board, and run the last round of benchmarks over an hour in DDR2-1050 mode, the G.skill DIMMs seem to have benefitted most from this extra airflow. Anyway, the perceived heat difference was noticeable, but not substantial - still, worth the couple of dollars that a memory cooler costs, compared to the DIMM price.

By the way, copper-coloured GEIL memory had the slight advantage of being colour matched with the Zalman heatsink, as you can see on the photo.

Benchmark Results
As the focus wasn't on the benchmark runs, but overclocking endurance and reliable work, only Sandra and 3D Mark 2006 were run - here is the sheet:

As you can see, Sandra memory bandwidth test performed somewhat better on the Nforce 680i vs 975X on the same clock and other matching settings. So, Nvidia memory controller seems to slightly beat the Intel's old one in latency and throughput.

The choice is, as always, yours
In the end, both 975X and 680i showed out well - the 975X somewhat above the expectations, while the 680i could do better. MSI needs to provide more comprehensive overclocking BIOS options, especially providing for lower multipliers in locked Conroes like ECS does. ECS, on the other hand, needs to cool the North Bridge better - either a better copper cooler with more efficient fan, or, well, water cooling provision without ripping apart those heat pipes.

The CPU match was a great disappointment with the X6800, but an equally great delight with the E6700 - all else being same, the cheaper flavour proved to be, in my case at least, a stunner. It completely outclassed the vastly more expensive Extremer - I'd like to hear from the readers how were their comparative experiences with these two CPUs? Maybe there is a pattern, after all: you can't know what you get inside that retail box... and, new steppings might be worth the try.

The memories have had pretty much the same performance in both in-sync and high-bandwidth modes, most of it having probably to do with their similar Micron die series, after all. In this round, I didn't go all the way to push them till the very end beyond the DDR2-1100 (out of the three DIMMs, only the GEIL part is certified by the vendor for that, up to DDR2-1160!), at least not until I get another Intel CPU with better FSB overclocking performance. I also believe that the DDR3 time is coming - the lack of chipsets was the main hurdle by now, but this will change in the next few quarters.

In summary, the 975X still holds out well, as you can see with this MSI board (and similar Asus and Abit offerings that Fudo tested here last week). The Nvidia 680i has a plenty of potential to improve, this being the first series - keep in mind, though, that not only the RD600 is around the corner, even being the last Intel core logic from that vendor (therefore they might just decide to push 'pedal to the metal' one last time), but also Intel will be enriching our 2007 experience with a whole series of brand new chipsets, too. I hope these new entries won't be DRM infected - a fat hope probably?

As for the CPUs, watch out for more new steppings before the "official FSB 1333" parts come out next year, and, well, try to see what you're buying in that box in the meantime - a newer stepping should mean greater overclocking potential, usually. Since most shops won't let you test your chosen CPU & board & memory combo together, do try to get some 30 day return policy if a particular part proves to be below the expected "performance override" potential - you might just get more lucky with the replacement part. ?