This spring we reviewed the 500 series of NVIDIA nForce chipsets (for AMD processors) and also had time to write a number of articles about motherboards on it, which are popular and attractive. However, there were no opportunities to shed light on the same nForce 500 series, but for Intel processors. Vendors announced several motherboards, but they were not in a hurry to the market. Anyway, this series has just become a matter of the past, because NVIDIA announced the new nForce 600i ("i" - Intel; "a" - AMD) chipsets that will completely replace predecessors and will introduce somewhat modernized architecture.
nForce 680i SLI
First of all, let's remember NVIDIA's series building principles: the top product supports x16+x16 SLI; the high-end solution supports x8+x8 SLI; the middle-end one doesn't support SLI, but equals to the previous two in everything else; and the low-end has features pared down and no SLI support. Therefore, judging by the products offered (nForce 680i SLI, 650i SLI, 650i Ultra), any reader can imagine the approximate features of nForce 600i series. While the nForce 500 line offered only two solutions for Intel processors: nForce 590 SLI and nForce 570 SLI, which just slightly expanded the assortment comparing to the nForce4 SLI IE and was designed only to introduce SLI support officially absent in other similar chipsets.
Now NVIDIA decided to expand to the middle-end market as well, seriously betting on nForce 600i functionality and company's reputation of "enthusiast's friend". By the way, company stated that the primary reason to cancel nForce 500 for Intel was their weak overclocking capabilities. Obviously, Intel would hardly be replacing the entire product series because of that. As for the long-anticipated NVIDIA's integrated chipsets, Intel processors still have to do with solutions from Intel itself and ATI - the profitability of this market is noticeably lower and it's hard to compete with Intel that successfully promotes its integrated home and office platforms.
Now let's concentrate on the top-end nForce 680i SLI, as we had a chance to test a motherboard on this very chipset. What functionality does it offer?
The functional balance between bridges is traditional enough: Northbridge (SPP) supports processor, memory, one PCI Express x16 and several PCI Express x1, while the Southbridge (MCP) supports peripherals, another PCI Express x16 (for the SLI mode), and the third graphics card slot at x8. The feature summary is as follows:
From the angle of architecture there's obviously only one innovation: spreading PCI Express interfaces across bridges to add space for the third graphics accelerator. In other aspects nForce 680i SLI is nearly identical to nForce 590 SLI (for AMD processors), except for processor and memory support, of course. And from the nForce 590 SLI IE it differs only by the third PCIEx16 slot. Due to this we recommend you to return for information on Southbridge operation to the nForce 500 article that describes networking and hard drive features of the chipset. And in this article we'll focus on processo, memory and graphics card support.
Of course, having got rid of issues with high-clock-rate FSB of nForce 500 IE, NVIDIA in every way underlines the efficiency of that feature in nForce 600i: 680i SLI is claimed to support 1333 MHz FSB. Other products officially lack it, but no one prevents motherboard vendors from implementing it. In all other respects everything's all right, quad-core processors work perfectly on nForce 600i boards that Intel even used for its own presentations. But naturally nForce 680i SLI advertising is not about "humble" 1333 MHz. Aimed at enthusiasts, this chipset promises nice overclocking potential. NVIDIA achieved 500(2000) MHz on the reference board, while we achieved only 450(1800) MHz on our sample. The potential was obvious anyway.
The dual-channel memory controller supports both 533/667/800 MHz official and higher non-official clock rates. Thanks to QuickSync technology, memory clock rate can be set completely independently, not via the fixed FSB clock rate divisors. In general, this advantage is not that important (since memory bandwidth and FSB balance is perfectly achievable in the usual way, e.g. in Intel chipsets), but at extreme overclocking this allows to use cheap memory (like DDR2-800) and enables to fine-tune the board for specific memory modules keeping balance of minimal timings and maximal clock rate.
Two fully-fledged x16+x16 PCI Express x16 slots allow SLI operation with theoretically higher performance than that of x8+x8. On the other hand, these slots are provided by different bridges, so the interconnect bus introduces additional latencies. At least that's what ATI claims offering chipsets where Northbridge supports both x16 slots. But these are just words, because it's impossible to test SLI on an ATI-based board or run CrossFire on an NVIDIA-based board. Of course, speaking about overclocking we must mention the capability of setting PCI Express clock rate independently and the LinkBoost technology that automatically raises PCIEx16 and HT clock rates by 25% when you install certified NVIDIA cards.
Finally, the third graphics card slot for a physics accelerator (by NVIDIA). This trend is very popular now and no need to mention that ATI is going to offer similar products, though information on those is scarce. It's just clear that this PCIEx16 slot (operating at x8) supports usual expansion cards as well (but this, unfortunately, won't stop spreading of PCIEx1 slots).
The chipset has some more peculiarities. It's hard to say how nForce 600i heat dissipation reduced (or increased) after the transition to the 90nm process, but the chipset is very hot. Reference heatpipe cooling did its job even without a bundled fan at maximum overclocking (450 MHz bus clock!), but we had no time for a serious examination of stability. Visual and tactile inspection makes us recommend to use a fan during overclocking and to think seriously about internal enclosure airflow. By the way, overclocking of NVIDIA's boards could be more convenient with the multi-functional nTune suite that allows to directly change BIOS settings and adjust nearly all system values (voltage, operation clock rate) in real time in Windows. Unfortunately, in our case the reference board and nTune 5.5 refused to work together, but we hope this issue will be solved.
Summing up the nForce 680i SLI functionality and comparing it with i965 and i975X (there's only sense in comparing it with Intel's latest Core 2 chipsets), we must state that NVIDIA's solution is considerably preferable from the angle of multi-processor graphics solutions (though it doesn't support CrossFire unlike both Intel's chipsets). As for processor and memory support, here rivals are on a par with each other (equal key features and different proprietary technologies). nForce 680i SLI has 2 x SATA channels more than i975X and 1 x PATA channel more than i965. The only advantage of Intel's chipsets is the support for Matrix RAID, because RAID 0, 1, 0+1, and 5 are available to all three. nForce 680i SLI can also boast about dual Gigabit controllers and their functionality (see article on nForce 500), to which Intel can offer only a MAC controller of i965 that requires a specific PHY controller. Finally, NVIDIA's product has 2 x USB ports more than i975X.
Though NVIDIA seems to clearly outperform its rivals, you shouldn't forget that other nForce 600i chipsets will have less functionality, including only 4 x SATA ports vs. 6 (fortunately, PATA support will remain and will even be expanded to 4 devices.) The single network controller will also be pared-down - as well as graphics card support. So junior chipsets of 600i series will be on a par with or even below Intel's solutions.
Not to overload the report with details we decided to compare just two products: the novelty and the fastest Intel P965 based board. The latter doesn't yield to i955X/975X and it's also nearly the only rival for nForce 600i (though 680i SLI is positioned as a higher-end.) We will also we'll see how well the preset memory and processor modes of the novelty work, will take a look at operation of Corsair's DDR2 memory capable of operating at 1142 MHz. Let's start with a low-level memory controller examination in our RightMark Memory Analyzer.
Unlike i965, there's no performance drop, when nForce 680i SLI works with DDR2-533. Though since the theoretical memory and FSB bandwidths are equal, it's "on the verge". Nearly all modes are equal, confirming the excessive memory bandwidth of faster memory, including DDR2-1142. As you remember, NVIDIA chipsets for Intel feature dedicated data and address bus for each DIMM module. This enables to use 1T command rate along with the usual 2T. But, judging by memory read speed, we can see that affects performance only by a couple of percents. Speaking of how nForce 680i SLI compares to Intel P965, NVIDIA's solution wins about 11%.
No sense in considering the peak memory write speed (if we exclude processor cache influence), all configurations are produce equal results. But things get more interesting in real modes. nForce 680i SLI shows a slight boost of memory write speed at higher memory clock rate, but the dependence on latency is obvious, so DDR2-800(4-4-4) doesn't surpass DDR2-667(3-3-3) (absolute timings - 10 ns vs. 9 ns). And DDR2-1142(5-5-5) with 8.75 ns timings shoots ahead a bit. It's hardly impressive, so there's no sense in purchasing very expensive memory because of some 5%. But what's interesting, only in this mode NVIDIA's chipset is able to outperform Intel P965. Note that operation at 2T command rate reduces numbers only by 3%.
Memory latency examination turned out to be very interesting: take a look at fantastic results of nForce 680i SLI at pseudo-random reading! High time to remember about the DASP unit for hardware prefetch and caching in the chipset. Those familiar with RMMA testing algorithm might want to look at visual differences in memory read latencies: Intel P965 draws usual jaggies, each of which depends on processor multiplier (which affects the relative positions of read instructions of benchmark's instruction stream). And nForce 680i SLI draws almost a straight line, indicating there's no memory access as such. The absence of difference between 1T and 2T command rates is also indicative, while at random reading the latter suffers from a serious penalty lagging behind by 8%. Obviously, DASP 4.0 perfectly caches open memory page providing fast access to its data and giving such wonderful results under conditions of preferred reading from one page at pseudo-random access.
We'll check the relevance of obtained results in real applications, but we must state that at true random access NVIDIA's chipset looks very good (though it has poor low-level latency results in RMMA). In this mode different nForce 680i SLI configurations look more natural and expected: only the DDR2-533(3-3-3) lags behind most with absolute timings of 11.25 ns, while the DDR2-1142(5-5-5) is the leader again. Thus it stimulates some demand by "the most enthusiastic" enthusiasts. Anyway, people purchase this kind of memory not for default operation modes.
Well now, let's see how the newbie performs in real applications.
Alas, only 1% of advantage is not a reason to celebrate. And this class of applications is the best for revealing the difference between memory controllers. nForce 680i SLI is slightly faster than i965 in the same mode; the transition to 2T command rate introduces 2% performance drop; the latency is actually lowest with DDR2-667(3-3-3) and DDR2-800(4-4-4), while the DDR2-1142(5-5-5) is the performance champion. Either Fast Memory Access enables Intel's chipset to catch up in real applications, or DASP errors negate NVIDIA's advantage... Long before tests are over, it's becoming obvious that competitors are merely equal.
Of course, we ran all necessary benchmarks for each nForce 680i SLI operation mode. But, as we had expected, the difference between final scores turned out to be negligible, so we left only the most important on further charts. Add to this identical CPU performance on both motherboards (i.e. both almost perfectly set FSB clock rate to nominal), and you get identical video encoding performance, measured according to our open-source method.
We had some hopes for the SPECviewperf benchmark. Though we were not dealing with professional 3D accelerators, this benchmark had repeatedly proved its capability of revealing differences in graphic interface implementations, intensively pumping data to graphics card. And it didn't let us down :) It indicated nearly identical results in 5 subtests of 8 (in some of which i965 was slightly better), but also showed significant advantage of nForce 680i SLI in the remaining 3: +6%, +11%, and even +18%. Nevertheless, for a casual user these values mean the same as those of RMMA: not many people run professional 3D modelling suites on home gaming machines. So, will these differences be proved by real games?
Whichever chipset you support, they perform the same.
Just maybe in FarCry supporters of NVIDIA might see a percent of advantage. For everybody else the case is obvious and closed.
Enthusiasts got themselves an excellent toy with flexible overclocking capabilities. Others got a highly-functional chipset, boards on which won't be cheap (about $200-300), but competitive, since their prices won't surpass those of i975/965 based solutions. It will be even harder to choose, because nForce 600i series didn't introduce performance boosts in normal modes. However, besides full-speed SLI support, the opportunity of installing the third graphics card for physics acceleration makes this product unique throughout the entire platform market.
Shipment of nForce 680i SLI based boards are to start in November. Some vendors (e.g. eVGA and Biostar) will be selling reference models manufactured by Foxconn. But, of course, we'll also see original solutions from companies like ASUS, Gigabyte, etc. (By the way, ASUS's ROG series board promises unforgettable impressions of tracery heatpipes.) Models on junior nForce 600i modifications will go on sale later for $150 or so, which highly increases their chances to become popular.
Sergey Pikalov (firstname.lastname@example.org)
November 10, 2006
Write a comment below. No registration needed!