One of the biggest events of 2004 was an announcement that NVIDIA was finally admitted to the market of Intel processors. The companies finally signed an agreement on cross licensing, which had been anticipated since the release of the first nForce generation. This chipset caused a sensation as soon as it came out. It seemed logical that NVIDIA should exploit the success and offer a solution for the most part of the desktop market. But time was getting on, Intel announced new FSB frequencies in its processors (the company requires separate chipset manufacturing licenses for each of them), but NVIDIA nForce still stayed in the AMD domain.
Californian chipsets undoubtedly won the recognition of manufacturers and customers: nForce3 and nForce4 deliveries amounted for up to a half of the AMD64 chipset market by the fourth quarter of 2004. But the company management evasively mentioned manufacturing products for Intel as an extra headache without clear prospects. It goes without saying that the tone of comments got inverted after the cross licensing agreement was signed: at last we can make a lot of money and realize the dream of every Pentium 4 owner.
Intel's reasons are obvious. Top Intel processors are more or less on a par with Athlon 64/FX in gaming performance, provided one video card is used. But if you take advantage of the full AMD64 potential with nForce4 SLI and use two top NVIDIA video accelerators in SLI mode, such a system will be a champion in 3D in the majority of cases. Yep, it will cost dearly (in the direct sense of this word), but the sector of ultimatistic highest-performance solutions is very important for the market in general — it's actually an attribute of a prestigious trademark, a foot-hold for the advertisement and promotion of all processors. Intel has not been trying to compete for the laurels of the most effective (price/performance) Low End solution, but it will fight to the end in the top segment.
And what do our readers gain from the release of nForce4 SLI Intel Edition? That's the topic of our article today.
NVIDIA nForce4 SLI Intel Edition Chipset
NVIDIA released an entire series of nForce4 chipsets for AMD processors, but Intel fans are offered (for now?) just a top version — nForce4 SLI Intel Edition. Frankly speaking, this product name is not very good — too bulky and hard to distinguish (from nForce4 SLI). With your permission, we shall contract it to nForce4 SLI IE — let's hope NVIDIA will have no problems with using a Microsoft trademark :)
You don't even need to have a look at the chipset diagram, to get an idea of its features, because all nForce4 SLI functions have migrated to the new product without modifications, except for the two inevitable things: CPU bus and memory controller integration. Remember that memory controllers are integrated into AMD Athlon 64/FX processors, that's why the chipset doesn't have it. It significantly reduces Northbridge complexity and dimensions and even allows to unite the bridges into a single chip, as in nForce3/4. The Intel Edition product perforce has a dual channel memory controller (NVIDIA expectedly chose DDR2) and the company again returned to the practice of two bridges. So, nForce4 SLI IE offers the following features:
We are not going to review nForce4 SLI IE functionality that is not different from the original (in nForce4 SLI). You can read the review of nForce4 chipsets for AMD64. There is also no reason to describe the CPU bus parameters — we can only note that the 1066 MHz frequency is also supported by Intel 925XE and VIA PT894/Pro (still on paper). What concerns processors, only two Pentium 4 Extreme Edition CPUs support it.
Let's analyze the only new unit — the memory controller. In nForce and nForce2 this controller actually consisted of two independent 64-bit controllers with cross strapping. In the first place, that was convenient for simultaneous processing of requests from various devices — for example, CPU and GPU (at that time NVIDIA was busy promoting a chipset with integrated video). In case of nForce4 SLI IE, the primary task of the controller is to accelerate CPU access to memory data. An important innovation here is a dedicated data bus and addresses for each DIMM module — now the system can operate in 1T command rate mode. Due to the lower burst length, the NVIDIA chipset can fill the bus with memory regeneration commands more effectively.
DASP (prefetching and caching data) is represented in nForce4 SLI IE by the third generation, optimized for multi-threaded operation (it has been used in NVIDIA chipsets since nForce). And finally, the new QuickSync technology allows to clock FSB and the memory bus independently, which provides an opportunity to use faster memory and to increase frequencies of these buses during overclocking.
Let's bring out the object of our today's analysis. We are not going to analyze the SLI effect: it's clear that similar processors will provide similar effect. Our articles in the Video section contain an impressive number of test results for AMD64 SLI. Thus, our today's task is unassuming: find out a preliminary performance level for nForce4 SLI IE as a usual desktop chipset compared to its competitors. You can surely understand why it's preliminary: we are testing an engineering sample. Who are its competitors? This question is even easier to answer, because there is no point in comparing the top product from NVIDIA with low end chipsets for 800 MHz FSB — the prices and functions of these solutions are very different. By process of elimination we get only one competitor — Intel 925XE (represented by the fastest motherboard).
NVIDIA chipsets with Athlon XP outperform their competitors in equal conditions (memory and FSB frequency). But what concerns Intel processors, the fastest chipsets are still the ones from Intel, other things being equal (though competitors sometimes prevail due to their support of non-standard frequencies, etc). It's very interesting to see how memory controllers from Intel and NVIDIA fair in direct competition. As nForce4 SLI IE supports DDR2-667, let's see the effect of this memory with correspondingly increased timings. In our case, we managed to use a couple of Corsair XMS2 PC2-5400 modules (DDR2-675) with the 3-3-3-12 timings in DDR2-533 mode and with the 4-4-4-12 timings in DDR2-667 mode. We research the issue of performance drop in case of asymmetric channel load only in production motherboards. But today we'll just install two identical modules into the slots belonging to different channels. Well, let's proceed to memory performance analysis in the RightMark Memory Analyzer test package, developed by our programmers.
The first test knocks the wind out: the i925XE is almost 1.5 times as fast in reading from memory.
It writes to memory a tad faster as well. Information about the chipset architectures is not enough to explain this fact, so we just state the fact (it's specific for RMMA access algorithms) and note that with the above mentioned timings nForce4 SLI IE looks better with DDR2-533 than with DDR2-667.
Another interesting picture — [Pseudo Random] Memory Read Latency. The i925XE and the NVIDIA chipset are approximately on a par here, the latter is at a small advantage. It's clear that DDR2-667 with such timings will always be slower in the latency tests (the operating frequency is higher by 667/533=1.25 times; the main timings — by 4/3=1.33 times), so you shouldn't pay too much attention to this mode.
But the epilogue is too fascinating: the random memory read latency in nForce4 SLI IE is approximately 5% lower than in the product from Intel. Remember that the random access differs from the pseudo random access in RMMA by faster D-TLB depletion (translation look-aside buffer), because it opens a lot of memory pages almost simultaneously. Certainly, the performance drop cannot be avoided in this case, but nForce4 suffers the least from this effect! To all appearances, we see the effect of DASP (3.0) — prefetching and caching data from memory.
Low-level memory tests are interesting but they don't generally provide an idea how the test object fairs in real applications. So, let's proceed to practical tests.
Archiving performance depends much on memory latency, but still we expected a noticeable advantage of the Intel chipset. But in fact the high memory read/write speed of the i925XE had no effect, it's on a par with nForce4 SLI IE. And the DDR2-667 mode, which has a worse latency, is outperformed by a couple percents.
Video encoding performance does not depend much on memory performance, in our case all the three systems demonstrated the same result. We can clearly see already that the i925XE has no performance advantage in real applications, as we could have thought from RMMA tests.
Of course, a processor and accelerator are critical for SPECviewperf (a benchmark for professional OpenGL application performance), but we have also seen a noticeable difference between memory controllers in various chipsets. This is not the case though, the results are similar, and nForce4 SLI IE is the winner. As this chipset outperforms the i925XE even with DDR2-667, we can assume that NVIDIA traditionally optimizes chipset operations (whether in drivers or in BIOS settings) with a video card.
Games demonstrate their own preferences. Doom 3 is a tad slow in low modes on nForce4 with DDR2-667, but it demonstrates no difference between the other two configurations.
Unreal Tournament 2004 is a tad faster on the i925XE, paying no attention to the screen resolution and graphics settings.
nForce4 shoots a couple percents ahead in FarCry at low resolutions, but already at 800x600 the breakaway between the contenders goes down to the inaccuracy of measurement level. Thus Intel and NVIDIA chipsets are generally on a par in games and other real tests, but there is some difference in 3D applications. We have already noted the possible reason, it can also be the effect of the SLI unit in nForce4.
…and loses nothing at the least — going back to the article title. Concerning the results of our today's tests, nForce4 SLI IE and the i925XE demonstrate the same performance. As we carried out our tests with the same timings on both systems (with DDR2-533), you can hardly expect any differences between these chipsets. nForce4 has a potential trump card (DDR2-667 support). But in order to play this card, it requires very fast (=expensive) memory, which can operate at 667 MHz with the main timings being about 3. Otherwise, you will get the results similar to those we have obtained today. We can also hope that the production motherboards will be faster, but the same thing concerned nForce…
But what's the point in counting percent fractions, when nForce4 IE features SLI! Read the articles in the Video section of our web site to get more information about the SLI effect in the games you are interested in. Resource-intensive state-of-the-art games gain much from the second accelerator in the majority of cases. It's quite clear that such a system will heavily outperform the i925XE as well as any other solutions currently available on the market.
Speaking of functionality, nForce4 SLI IE has a lot to boast of. Its characteristics are perhaps the best among all solutions for modern Pentium 4. No one promises a free lunch though, prices for motherboards on this chipset will hardly please regular users (the leading manufacturers are expected to offers their products for over $200). But this product is not intended for regular users, few of them will have Pentium 4 Extreme Edition and two top NVIDIA accelerators. So, without taking into account the overall price of a system, nForce4 SLI Intel Edition is currently the best solution. Considering is price, it's a match for the best chipsets for Socket 775.
Write a comment below. No registration needed!