Until the most recent times choosing a platform for AMD Athlon CPUs never aroused any questions. In fact, since the arrival of Socket A processors and till the end of 2000 there existed only one core logic to support Athlon and Duron CPUs - the famous VIA KT133. AMD also had due resources for chipset development and production, but the company preferred not to pop onto the Socket A chipset market giving green light to VIA. Subsequently, throughout 2000 nearly all the Socket A mainboards for Athlon CPUs were based on VIA KT133. It left the buyers no alternative: they were destined to choose between different mainboards based on one and the same core logic.

However, this state of things upset both the users and manufacturers. First of all, these were the users who were unhappy with either the performance or the stability of VIA KT133, since they could buy nothing else but systems with Intel CPUs. Then, the poor choice of chipsets for Socket A CPUs barred the advance of AMD's processors to the market too.

Today we can observe a dramatic twist of the situation. One of the triggers was the coming of a new DDR SDRAM memory type. The necessity to implement the support for the new memory technology creted more or less equal conditions for the chipset manufacturers. So, that those companies, which hadn't ever offered any Socket A chipset solutions, decided to enter this market having no fear of the mighty rivals with a serious background. However, the old VIA KT133 core logic was kicked out of the market at once when it became obligatory for the modern Socket A chipsets to support EV6 CPU bus. It works at 266MHz and is the privilege of some high-performance Athlon CPUs. This concurrence resulted in an appreciably vaster range of Socket A chipsets. Now there are four manufacturers in this field: ALi, AMD, SiS and VIA.

As soon as we got an opportunity to choose, it's high time we compared the available alternatives. In this investigation we'll merely compare all the up-to-date chipsets for Socket A processors intended for High-End and mainstream systems. From our point of view, the basic criterion for the chipsets to be called "up-to-date" is the support of CPUs with 266MHz bus, which is also a must for the chipsets intended to support the upcoming Athlon processors on the new Palomino core. Another feature proving that a chipset is not targeted for the Low-End systems is the absence of an integrated graphics core. At the same time, we decided not to distinguish between the chipsets supporting DDR SDRAM and not. The price for PC2100 DDR SDRAM modules has plunged deep enough for us to believe it is reasonable to test DDR and non-DDR chipsets together. Moreover, it enables us to assess the real performance gain granted by DDR SDRAM.

So, finally we selected the following 5 chipsets to take part in the today's race:

• ALi MAGiK 1
• AMD 760
• SiS 735
• VIA KT133A
• VIA KT266

We would like to stress that apart from mainboards built on these five chipsets and selling worldwide now, in about a month we'll see one more DDR Socket A chipset, NVIDIA nForce. But as there are no ultimate samples of nForce based mainboards yet, we didn't include it in this comparison. Indeed, NVIDIA's offspring is out and away a different product and it would make sense to make a special report about it. We promise to do it some time later. Thanks to its dual-channel DDR SRDAM controller and DASP (Dynamic Adaptive Speculative Pre-Processor) technology, nForce has a fair chance to be the fastest core logic for AMD Athlon platforms. However, since it features a complicated built-in graphics core and audio-processor, the mainboards based on nForce will turn out a lot more expensive than other boards based on "conventional" Socket A chipsets. That's why we'd better review NVIDIA nForce as an integrated chipset for the mainstream and even High-End sector, but not as a competitor to discrete chipsets for Athlon CPUs.

Before we pass over to analyzing the results of the tests, let us highlight the peculiarities of all the chipsets participating and single out the unique features typical of each of them.

Testing Participants

AMD 760

AMD 760 core logic is the very first DDR chipset for Socket A processors. Although ALi officially announced its MAGiK 1 a little bit earlier, AMD 760 became the first DDR chipset demonstrated in public. It was introduced together with Athlon CPUs with 266MHz bus. Originally, AMD didn't intend to position itself as a chipset manufacturer. Therefore, AMD 760 just like its predecessor, AMD 750, was to serve solely as a means of promoting DDR SDRAM technology. AMD didn't feel like shipping AMD 760 in mass for a long time and looked for the opportunity to cease the production of AMD 760 as soon as possible, giving way to the DDR chipsets for Socket A by other manufacturers.

For this reason the abilities of AMD 760 are not as impressive as those of other chipsets. Besides, it's somewhat troublesome to produce mainboards based on this core logic.

AMD 760 comprises AMD-761 North Bridge and AMD-766 South Bridge traditionally connected via the PCI bus. Since AMD didn't want to make AMD 760 a mass product, the developers didn't focus on the performance of the devices connected to the South Bridge and, unlike other chipset makers, implemented no special bus faster than PCI to connect the Bridges. In theory, AMD could have done it. Take AMD 760MPX, a version of AMD 760 with dual-processor support. In this core logic AMD implemented 66MHz PCI64 to connect the Bridges.

The second restriction in AMD 760 is the strictly synchronous work of the memory and FSB. Therefore, processors with 266MHz bus working on AMD 760 mainboards require PC2100 DDR SDRAM, whereas CPUs with 200MHz bus limit the memory clocking to 100MHz.

One more inconvenience worth mentioning when we talk about AMD 760 is the support of only four memory banks when common, not registered DDR SRDAM is used. In practice it implies that on most AMD 760 based mainboards there are just two DDR DIMM slots. Worse still, AMD 760 supports only DDR SRDAM and doesn't work with PC133 SDRAM at all, being the least flexible DDR Socket A core logic of all.

The South Bridge, AMD-766, doesn't please us with a variety of functions either. Unlike most other South Bridges, this chip doesn't have integrated support of AC'97 sound and modem, has no integrated LAN controller and supports only four USB ports.

Surprisingly enough, AMD 760 is still popular. Until recently, other chipset manufacturers couldn't offer anything comparable to AMD 760 in performance, so AMD changed its mind and thought: why not? So, AMD 760 shipments kept coming.

It's for the splendid performance that AMD 760 retains its popularity among the mainboard manufacturers. For the sake of the chipset's high performance, they put up with its numerous shortcomings. Apart from the above mentioned issues, AMD 760 based mainboards are more expensive to produce than mainboards based on other chipsets: firstly, AMD 760 itself costs more than the rivalry chipsets, and secondly, it requires six-layer PCBs which are also more pricy than the regularly used four-layer ones.

Time passed and the mainboard makers contrived to settle some of these problems. Say, instead of the AMD-766 South Bridge, they use a cheaper and smarter VIA 686B. And the clever guys from ABIT, Chaintech and EPoX have improved the reference design of AMD 760 based mainboards and now they use four-layer PCBs instead of the six-layer ones, having reduced the costs a little bit. As a result, in spite of AMD's not very far-reaching plans and far not the best features of the chipset, AMD 760 is still in demand on the market.

ALi MAGiK 1

Though ALi MAGiK 1 was announced earlier than AMD 760, the arrival of real mainboards based on it took place almost together with those based on the latter. For quite a long period of time they were the only Socket A chipsets with DDR SRDAM support. AMD 760, as we have already said, boasted higher performance, while ALi MAGiK 1 won its market share mainly due to its rich features set and low price. Let's discuss it somewhat more thoroughly.

ALi MAGiK 1 consists of M1647 North Bridge and M1535D+ South Bridge connected by the PCI bus, like the bridges of AMD 760. It looks as though ALi didn't master any faster technology for inter-bridge connection when MAGiK 1 was developed. In its upcoming core logic for AMD Athlon platform, which is expected to come out in the first half of 2002, ALi is going to eliminate this bottleneck and replace the PCI bus with AMD HyperTransport technology.

The key peculiarity of ALi MAGiK 1 that allowed this core logic to compete with AMD 760 was its memory configuration flexibility. The chipset supports up to 3GB DDR SDRAM as well as the common PC133 SDRAM. This gave the mainboard manufacturers an opportunity to create boards supporting both memory types. In the beginning of DDR SDRAM onslaught to the market, solutions of the kind were really welcome since they enabled the users to gradually migrate from one memory type to the other. Furthermore, ALi MAGiK 1 allows asynchronous memory and FSB clocking, but most manufacturers disabled this function because of the utterly low performance this core logic showed in asynchronous mode.

Unfortunately, low performance became the main drawback of ALi MAGiK 1, which hampered the sales. Independent tests also run in our lab showed that MAGiK 1 didn't work much faster with PC2100 DDR SDRAM than VIA KT133A worked with PC133 SDRAM. In some tests ALi MAGiK 1 was even slower than the PC133 SDRAM chipset from VIA.

Attempting to stimulate the sales, ALi resorted to the tried-and-true method of price cuts. It led to a dramatic price gap between AMD 760 and ALi MAGiK 1 (the latter turned nearly two times cheaper than its competitor!), and of course, it couldn't but tell on the ultimate price of the mainboards. The buyers were often fooled with a colorful "DDR" marking on a box with a surprisingly cheap mainboard: DDR SDRAM eventually gave no performance gain with a Socket A mainboard built on ALi MAGiK 1…

However, it was too early to write off ALi MAGiK 1. Shortly after the poor performance of MAGiK 1 was revealed, ALi promised to develop a new chipset revision where this bug would be removed. Half a year later the revision came into being.

It is numbered as "B0"and slightly differs from the previous "A0" in marking:

Old Revision - A0		New Revision - B0

The new revision of ALi MAGiK 1 has a notably improved memory controller, which has contributed a lot to its performance. In order to estimate the performance growth of B0, we compared the performance of ALi MAGiK 1 A0 and B0 in several most popular tests.

The performance gap in office and content-creation applications is 4-5%. Well, it's still better than nothing, isn't it?

The results B0 shown in Quake3 Arena are just amazing: the performance gain makes 34%! The new revision of ALi MAGiK 1 seems not only to have a higher memory bus bandwidth, but also lower latency, because this characteristic is crucial in Quake3 Arena.

In Serious Sam the performance upturn is not as outrageous as in Quake3 Arena, but even 13% look really great.

As a whole, we admit that the new revision of ALi MAGiK 1 is much faster than the old A0 revision. Now ALi MAGiK 1 is fit enough to compete with other DDR chipsets for Socket A processors.

Sadly, this news doesn't rejoice the owners of mainboards based on the old A0 revision. In case you're only planning to buy a mainboard based on ALi MAGiK 1, please, be careful not to overlook the number of the chipset revision.

VIA KT133A

Almost together with AMD 760 and ALi MAGiK 1 supporting DDR SDRAM and the new Athlon CPUs with 266MHz bus, VIA introduced a chipset of its own, VIA KT133A. It was also designed for this kind of Athlon CPUs, but supported only PC133 SDRAM. VIA KT133A came right in time: late in 2000 and early in 2001 DDR SDRAM was far more expensive than PC133 SDRAM modules, and the accelerated bus of the new Athlon CPUs could give them a tremendous performance boost without DDR SRDAM. So, the "mid-way" VIA KT133A fell in fat soil.

As a matter of fact, VIA KT133A is a slight recast of the well-known VIA KT133, a chipset which was so much beloved by the Socket A systems manufacturers. The new modification has a North Bridge with corrected internal timings, so now VIA KT133A is able to easily increase the FSB frequency up to 150-160MHz and officially supports 133MHz FSB. Just for a better comparison, the old KT133 endured nothing more than 110-115MHz FSB and the system simply hung when the FSB frequency got higher than that.

VIA KT133A doesn't feature any architectural innovations. Like KT133, it comprises VT8363A North Bridge and VIA 686B South Bridge connected with the PCI bus. The core logic supports up to 1.5GB PC133 SDRAM and allows clocking the memory and CPU buses asynchronously.

The most favorable price-to-performance ratio of KT133A + PC133 systems made VIA KT133A a bestseller among all the other chipsets for Athlon CPUs with 266MHz bus, although they boasted DDR SDRAM support. The well-being of VIA KT133A was rocked by an annoying bug detected one day. The trouble was that while copying huge files from one IDE device to another connected to different IDE channels, the data was sometimes distorted. Very soon the new BIOS versions and IDE Bus Master drivers from VIA got rid of the bug, but this slip undermined the public's trust to KT133A. Nonetheless, VIA KT133A stays popular with those who want to build a low-cost system on Athlon or Duron CPUs.

VIA KT266

Knowing that AMD 760 and ALi MAGiK 1 had a number of shortcomings (the first one was too expensive and couldn't boast rich features set, while the performance of the latter left much to desired), VIA pinned its faith on a new DDR Socket A core logic, VIA KT266. After the launching of VIA KT266, expected to become a real king of the market, AMD could supposedly stop shipping its "temporary" AMD 760. At least, the promising characteristics of VIA KT266 gave every reason to think so.

The memory controller of VIA KT266 proved smarter than those of ALi MAGiK 1 and AMD 760. VIA KT266 supported both key memory types (DDR SDRAM and PC133 SDRAM) and allowed asynchronous clocking of the memory and CPU buses. It also supported up to 3GB unregistered DDR SDRAM (up to 4GB in case of registered modules). This way, both VIA KT266 and ALi MAGiK 1 let the manufacturers create mainboards with DDR SDRAM and PC133 SDRAM support and the maximal memory size.

Another bunch of functions is granted to VIA KT266 by its new South Bridge. It has an integrated LAN controller and six-channel AC'97 sound that are not implemented on any other Socket A core logic. The Bridges (VT8366 North and VT8233 South Bridge) are connected by means of a special V-Link with 166MB/sec bandwidth, i.e., it's twice as fast as the notorious PCI bus of the above mentioned chipsets. This way, VIA aimed at guaranteeing that none of the devices implemented in the South Bridge would lack the bandwidth of the bus connecting the memory and the CPU.

With all this variety of functions, VIA KT266 was pricing at a reasonable level and didn't require six-layer mainboard design. At first glance, VIA KT266 couldn't fail to win the affection of the mainboard makers and ultimate users. But...

But the tests following the launching of VIA KT266 mainboards showed that they were, of course, faster than those boards based on the first revision of ALi MAGiK 1 chipset and those built on VIA KT133A, but they couldn't catch up with AMD 760 based boards. A little bit later AMD announced its intention to go on shipping AMD 760, because it was yet displeased with the performance of VIA KT266, whereas AMD 760 provided the highest performance for all Athlon systems. VIA promised to develop a new revision of its KT266 and to improve the memory controller. In a while VIA intentions simply vanished and the new chipset revision was left up in the air. No new revision appeared when the due date came. Moreover, the company's officials didn't say a word on the R&D work in this direction. Apparently, it was not that easy to improve KT266. in other words, a faster chipset version either won't come out so soon, just the same way it happened with ALi MAGiK 1, or VIA has simply given up the mere idea of making any modifications and won't move a finger to change anything in its KT266.

Anyway, VIA KT266 still has quite a promising future ahead. This core logic features richer functions than its counterparts and it's a good basis for cheap DDR mainboards. Besides, VIA's warm relations with the Taiwanese mainboard manufacturers are sure to create fertile climate for KT266 and to make mainboards built on it go off like hot cakes.

SiS 735

The last participant involved in our roundup of the up-to-date Socket A chipsets is SiS 735. This is the company's first attempt to regain its positions in the high-performance core logic market. SiS 735 isn't too late to come into the game, though it is launched later than the other chipsets: DDR SDRAM prices have just fallen to an acceptable level, so a new wave of interest to DDR SRDAM solutions is to arise in the months to come. Another favorable factor for SiS 735 is that all the other DDR Socket A chipsets didn't prove up to everybody's expectations. This time even the traditionally skeptical attitude to SiS chipsets will hardly prevent SiS 735 from getting popular. It really deserves keen attention.

SiS 735 has two bright peculiarities. One of them will catch your eye as soon as you take a look at any mainboard built on this core logic, the other one is hidden deep in the chipset. The first "unusualness" is that SiS 735 consists of only one chip containing both the North and South Bridges. This 2-in-1 solution allows reducing the costs of the chipset itself and mainboards based on it, because SiS 735 requires not so big PCBs. The other peculiarity is a specially designed bus connecting the Bridges integrated into a single microchip. It is known as MuTIOL (Multi-Threaded I/O Link) and has 1.2GB/sec bandwidth. But the gigantic bandwidth is not the only unique trait of this bus. Its main advantage is competitive access guaranteed to every device connected to the bus. That is, since every device is given an individual parallel stream, the latency gets much lower and scarcely any collisions occur. When the inter-bridge bus is heavily loaded, MuTIOL transfers the data between the CPU, memory and devices connected to the South Bridge of SiS 735 considerably faster than PCI or V-Link buses implemented in the other chipsets.

As for the memory controller, SiS 735 is not equipped with anything special. It supports the same functions as the controllers of ALi MAGiK1 and VIA KT266: DDR SDRAM plus PC133 SDRAM support and also the possibility of asynchronous memory and CPU bus clocking. The maximum memory supported by SiS 735 makes 1.5GB.

All these features are surely attractive, but the main advantage of SiS 735 is its price. According to SiS plans, SiS 735 will become the cheapest DDR Socket A chipset in the High-End sector. Combined with the appealing characteristics, it will help SiS 735 find its way to the hearts of both the manufacturers and users. On the other hand, SiS risks being troubled with its ancient problem - the lack of production capacity. In the long run, SiS may turn out unable to satisfy in full the demand for SiS 735 chipsets. That's why most mainboard makers don't venture to deal with this company. Fortunately, there are such exceptions as ASUS and ECS that have included SiS 735 based mainboards in their nearest plans.

Chipset Specs Comparison

In this table we've summarized the characteristics of all these five chipsets considered:

DDR SDRAM

The last topic we'd like to address before reporting the obtained test results is the much talked-about DDR SDRAM, one of the main characters of this article.

The fact that the shift to DDR SDRAM production required trifling technological changes has finally made the right influence on the memory makers. Besides, many new mainboards with DDR SDRAM support started cropping up in the market, which also couldn't remain unnoticed. Consequently, most of the memory manufacturers got encouraged enough to start producing DDR SDRAM in mass offering PC2100 DDR SDRAM modules together with PC133 SDRAM ones. Nowadays however, the new modules are a common thing to find in the stores.

We mean not only 200MHz PC1600 DDR SDRAM, but also the fully fledged PC2100 DDR SDRAM, which is famous for its better performance and works at 266MHz. Further technological improvement enabled the leading memory makers to ship PC2100 with CAS2 instead of CAS2.5. For DDR SDRAM this mattered a lot more than for PC133 SDRAM. Reducing CAS Latency from 3 to 2 in systems with PC133 SDRAM brings about a negligible performance gain, while in case of systems with DDR SDRAM this change influences the performance more discernibly. The trick is that the memory controllers of some DDR chipsets can't work properly with the non-integral CAS Latency equal to 2.5 and use CAS Latency equal to 3 instead. That is why for DDR SDRAM systems PC2100 CAS2 is certainly preferable.

Memory chips with CAS2 have been manufactured by the major brands like Micron and Nanya for quite a long time now, therefore it's pretty easy to find memory modules based on these chips. Most of CAS2.5 chips feel well at the lowest timings including CAS Latency equal to 2. For example, look at the CAS2.5 PC2100 module we tested in our lab - it works no worse than the original CAS2 memory.

It was for these particular reasons that we decided to test our DDR platforms with CAS2 memory. This way, the DDR platforms found themselves on equal footing with PC133 SDRAM platforms featuring CAS2 memory since quite long ago.

Testbed and Methods

So that to compare the systems built on different chipsets in a due way, we did our best to squeeze all juices from the tested samples. First of all, it has to do with the BIOS settings and the memory. As we have already mentioned, we used PC2100 CAS2 DDR SDRAM in DDR systems and PC133 CAS2 SDRAM in the PC133 SDRAM system. We set all the timings in the BIOS Setup to the maximum and enabled the chipsets specific functions serving to improve the performance (Bank Interleaving, Super Bypass).

Moreover, we engaged today's most powerful CPU and graphics card - AMD Athlon 1.4GHz with 266MHz bus and a card built on NVIDIA GeForce3 graphics chip. So, we can regard the obtained results as the top performance one can get nowadays on systems like those.

We selected the following mainboards built on various chipsets: ABIT KT7A (one of the fastest KT133A based mainboards), MSI K7T266 Pro (the only certified KT266 mainboard from AMD so far), EPoX EP-8K7A (the best AMD 760 based mainboard of the latest generation), the reference mainboard based on SiS 735 and the one on the new revision of ALi MAGiK 1.

That's what our test systems looked like that:

In office and gaming applications we tested in Windows 98SE. Tests in the professional OpenGL applications were run in Windows 2000 Professional SP2.

Performance

First, let us see how things stand in the synthetic tests indicating the performance of the chipsets memory subsystems. These results will give us an idea of their abilities in real applications.

When we compared the figures obtained during this test session and those from our previous tests, we had to admit that the two newcomers (the new revision of ALi MAGiK 1 and SiS 735) have a lot to be proud of. We should also call your attention to the new BIOS versions that have affected the performance of VIA KT266. Surprising as it might seem, the leader in SiSoft Sandra is SiS 735. The former leader among DDR chipsets, AMD 760, runs the second with the engaged FPU or even the fourth if we consider the bandwidth indicated during operations with integer numbers.

Cachemem benchmark casts more light on the memory controllers implemented in different chipsets. As we can see on the diagrams, the improved ALi MAGiK 1 is now the fastest in reading from the memory, but still fails to be the first in writing and shows upsettingly big latency. Nevertheless, as long as reading from the memory is needed more often than writing, MAGiK 1 will be good in real tests. The newly created core logic from SiS, SiS 735, is fast at reading, too. Considering the lowest latency SiS 735 reveals, we can acknowledge it as a DDR chipset with the most effective memory controller. The former performance leader, AMD 760, can boast only extremely fast writing into the memory, and VIA KT266 is far from the top everywhere. The developers of some DDR chipsets achieved unexpectedly low latency in their products. It is even lower than that of PC133 SDRAM working with VIA KT133A. This fact shows that VIA feels pretty relaxed as there is no real competition and doesn't take the pains to optimize its Socket A PC133 SDRAM chipset. As for the Socket A chipsets with DDR SDRAM support, there are four of them, so each of the developers did its best to squeeze the most out of his offspring. As a result, we dare conclude that a the healthy competition in the Socket A chipset market told really positively on the entire situation.

Now let's check how close the results of synthetic tests are to those shown in real applications.

The new version of this Bapco test helps to evaluate the performance in the most popular real office and content-creation applications. Please, mind the fact that the new SYSmark 2001 test doesn't measure the performance in separately run applications providing some average result in the end, but emulates real working conditions launching the tasks parallel to each other and switching between them. The way the chipset works with the memory, as well as the performance of the disk subsystem, which in its turn depends on the quality of the IDE Bus Master driver, tell a lot on the final result of the test.

In this benchmark SiS 735 is at the top of the list thanks to its fast memory and the fastest bus for inter-bridge connection. ALi MAGiK 1 represented by the new revision is no outsider any more but enjoys the second position in the leading group. The comparatively poor performance of AMD 760 is explained by the widely-used combination of AMD-761 North Bridge and VIA 686B South Bridge being the slowest in working with the disk subsystem, as we have already mentioned in our previous reviews.

The main idea of Internet Content Creation test in SYSmark 2001 is to measure how fast the systems can create a real web-site. The following applications are involved in this process: Adobe Photoshop 6.0, Adobe Premiere 6.0, Macromedia Dreamweaver 4, Macromedia Flash 5 and Microsoft Windows Media Encoder 7. The fastest chip in this contest appears VIA KT266.

Office Productivity test emulates typical office work making use of such applications as Microsoft Word 2000, Microsoft Excel 2000, Microsoft Power Point 2000, Microsoft Access 2000, Microsoft Outlook 2000, Netscape Communicator 6.0, Dragon NaturallySpeaking Preferred v.5, WinZip 8.0 and McAfee VirusScan 5.13. We were not surprised to see the leaders SiS 735 and the new revision of ALi MAGiK 1 among the leaders. It's a curious thing that VIA KT133A with PC133 SDRAM doesn't look bad at all running alongside with the DDR chipsets. It stamps DDR SDRAM platforms as not necessary for office applications.

Business Winstone 2001 test deals with typical office applications too, but in this case memory latency and the disk subsystem abilities play an even more decisive role. The results are quite natural: SiS 735 and VIA KT133A head the race, proving once again that in regular business applications you can get by without increasing the memory bus bandwidth at all.

The line-up alters when we pass over to a more memory-demanding test. In content-creation applications SiS 735 takes the advantage of its fastest DDR SDRAM controller and breaks ahead.

It's important to say that different results shown in Winstone and SYSmark can be explained by the fact that the tests use different scripts. Besides, the Winstone benchmark set is a bit older than SYSmark and hence uses slightly older versions of the applications included into it.

To make the office coverage complete, we measured how fast the systems could archive the data packs of big size (directories with installed Unreal Tournament). For this purpose we used a widely-spread WinZIP archiving utility. Apart from constant memory addressing, this program was continuously working with the disk subsystem. That is why the results of this test let us not only see how efficient the buses between the Bridges are but also evaluate the quality of IDE Bus Master drivers implementation. On the diagram above you can see the time spent by each system on archiving the same directory with UT. The shortest time stands for the best performance.

Well, the situation is not surprising at all. The worst results belong to those chipsets with PCI bus connecting the bridges: they are AMD 760, ALi MAGiK 1 and VIA KT133A. However, the well-optimized Bus Master driver of KT133A saves it from a complete failure. The indisputable leader here, SiS 735, owes its success to a new MuTIOL bus.

At last, it's time to see what the contemporary Athlon chipsets are worth in games.

Quake3 brings about pretty illustrative results. The performances of various DDR chipsets differ a lot, not to mention the gap between the DDR group and the PC133 core logic. The diagram shows that the performance in Quake3 depends directly on the read rate from the memory. Therefore no wonder that SiS 735 and the new revision of ALi MAGiK 1 break ahead of the other three racers. They are 14% faster than the PC133 SDRAM platform and almost 10% faster than AMD 760, a present performance paragon of all the DDR Socket A systems. As for AMD 760, it and VIA KT266 go abreast overtaking VIA KT133A with PC133 SDRAM by feeble 4%. The resume is that the alliance of DDR SDRAM and Athlon CPUs is fruitful only if the memory is combined with a "due" chipset.

Higher resolutions in Quake3 level out the results, though even in 1280x1024x32 mode, when the graphics accelerator works the hardest of all, SiS 735 and ALi MAGiK 1 don't loose their dominance.

The same tendency exists in Unreal Tournament. The fastest is SiS 735, then follows ALi MAGiK 1. VIA KT266 is awfully slow: its performance is only 2% higher than that of VIA KT133A with PC133 SDRAM.

At 1024x768x32 ALi MAGiK 1 rev. B0 overruns SiS 735 and occupies the leading position making use of its unbelievably fast reading. All other results repeat the previous case pretty much.

We have also tested the systems in Serious Sam. It's clear from the diagram that the chipsets behave in the same manner as they did in Unreal Tournament. The fastest DDR SDRAM platform, the one with ALi MAGiK 1, overtakes VIA KT133A with PC133 SDRAM by good 12%.

VIA KT266 has disappointed us again. As the resolution in Serious Sam grows, this core logic becomes even slower than its PC133 SDRAM brother, VIA KT133A.

MDK2 is a relatively old gaming application, which is sensible to the memory bandwidth while reading. It puts ALi MAGiK 1 out of competition - it is a real winner. AMD 760 and VIA KT266 with DDR SDRAM have nothing to boast in this game.

Like in all the other games, increasing the resolution leads to the general figures leveling.

3DMark 2001 gives us a chance to see the chipsets performance in the upcoming games with DirectX 8.

The situation here is similar to the other games. ALi MAGiK 1 and SiS 735 retain the fame of the fastest Athlon chipsets (especially in the "Lobby" game). In the hardest "Nature", however, all the chipsets show similar results. A typical feature of this test is the working hard graphics accelerator, so the memory bandwidth and latency are not so important. These results point that caring about the best core logic yields the highest performance only if the entire system is well-balanced. Otherwise, some slower devices will prevent the rest of the system from showing its best.

No decisive changes occur in 3DMark 2001 at higher resolutions. We should only do justice to SiS 735 leaving ALi MAGiK 1 behind in most of the tests. For some reason, ALi MAGiK 1 system hung at Dragothic test and didn't pass it. Obviously, the new revision of MAGiK 1 hasn't been completely finalized yet and some bugs are still there. Indeed, in course of our tests we discovered one more problem with B0 revision of ALi MAGiK 1. The mainboard based on this core logic lost stability when dual-bank DDR SDRAM modules were used. Hopefully, the guys from ALi will soon settle all these problems.

The last test we ran was SPECViewPerf emulating the work of professional OpenGL applications. The apps of the kind are utterly sensible to the real memory bus bandwidth, so we were particularly interested in the results.

We have already stressed that the performance in SPECViewPerf is mainly determined by the memory bandwidth, not the latency. These are the best conditions for ALi MAGiK 1 with its fast reading from the memory and AMD 760, which is the leader in writing. The other couple of DDR chipsets, SiS 735 and VIA KT266, couldn't get that high.

Conclusion

The time has come to give our verdict. To start with, we should say that the first generation of DDR SRDAM controllers that we based our conclusions about the performance of this memory type on, have revealed their inadequacy. New DDR memory controllers implemented in the latest chipsets prove that DDR SDRAM is more effective that it used to be thought. Besides, PC2100 DDR SRDAM with CAS Latency equal to 2, also contributes to the higher performance of DDR systems. The benchmarks show that the performance difference between Athlon platforms with DDR SRDAM and PC133 SDRAM ranges from 5-7% in office applications to 15% in games and 45% in professional OpenGL applications. These figures can be the crucial argument in favor of DDR SRDAM, especially since these modules have considerably dropped in price recently.

Analyzing the progress made by each of the five Athlon chipsets, we should state with surprise that the brand new product from SiS, SiS 735, and the new revision of the former outsider, ALi MAGiK 1, have turned out the most successful solutions. Their brilliant performance is out of the question, but the market outlook for these two chipsets is still uncertain. Both SiS and ALi can face a serious lack of production capacities and appear unable to satisfy a sudden excessive demand for their chipsets. Nonetheless, it doesn't prevent SiS from such daring actions like the recently expressed intention to occupy up to 50% of the chipset market by the end of the year. There is one more predicament for these chipsets: neither SiS nor ALi are greatly beloved by the mainboard makers and the end-users. None of these companies has made any high-performance product for a very long time, being brought into disrepute as a developer of second-rate solutions. It will take SiS and ALi a while to eliminate this deep-rooted attitude. No doubt, their chief trump will be pricing policy.

The previous A0 revision of ALi MAGiK 1 is already known as a relatively slow solution. Of course, the new B0 revision works much better, but the customers may still be very uneager to buy ALi MAGiK 1 based mainboards, because the only way to see which North Bridge is installed on the mainboard is to remove the chipset cooler.

Anyway, AMD has every reason to be happy: new chipsets with high-performance memory controllers which are to replace AMD 760 have finally come into being. ALi MAGiK 1 and SiS 735 are scheduled to start sipping in mass since August, and at that time AMD theoretically may give a sigh of relief and stop producing its AMD 760.

We have a special word to say about VIA KT266. Last time when we tested this core logic, we hadn't yet got hold of ALi MAGiK 1 and SiS 735, so VIA KT266 seemed quite competitive to us. It was close to AMD 760 in performance and enabled the manufacturers to make somewhat cheaper mainboards. Our opinion changed radically after this test session. Now it appears absolutely clear that until VIA launches a new revision of this DDR Socket A chipset (if it is going to launch any at all), KT266 will be a failure. Without the new revision, VIA can only hope for its contacts with many mainboard makers to whom KT266 will be palmed off.

Well, let's wait till fall when mainboards built on the new DDR chipsets for Athlon CPUs will start selling worldwide. It will provoke a real positional rearrangement among the Socket A chipset manufacturers. Time will show what this affair will end in. At this stage the most evident consequence of the new chipsets' arrival is a bunch of new solutions to strengthen the rating of both older and latest Athlon CPUs based on Palomino core.