Introduction

Once again we have to address the much talked-about topic of DDR chipsets for AMD Athlon CPUs. It looked as though we dotted all the "i"-s a month ago having written an article on comparing modern Athlon chipsets, but due to VIA this review has to be updated.

Today, on September 3, VIA has officially introduced a new revision of KT266, a good old DDR Socket A core logic. The new chipset is named KT266A:

Before we pass over to discussing the differences between the old KT266 and the newly born KT266A, let us remember why KT266 failed to gain broad popularity, though it's a pretty smart product. Indeed, if compared with other rivalry chipsets, the specifications of KT266 look nice (for details you're welcome to consult our article "Contemporary Socket A Chipset Comparison"). However, performance tests revealed the chipset's inability to live up to the users' expectations. Mainboards based on KT266 did not only had poorer performance than those based on the paragon AMD 760, but also fell far behind mainboards built on a much cheaper SiS 735. VIA did its best to inveigle the mainboard makers to use KT266, tempting them with all sorts of concessions, but none of these attempts was a success. Eventually, VIA KT266 didn't become a popular DDR chipset for Athlon CPUs, though the whole hardware community expected it would. Even AMD said no flattering words about KT266 performance.

All this fuss inspired VIA to go on working on the already launched KT266 in order to improve its performance. In the end of July the company promised to start manufacturing a new revision of KT266 with an improved memory controller. But there have been no talks about this revision up to now.

Anyway, it's better late than never: the new KT266 revision finally came into being. VIA has obviously decided to prevent any bitter reminiscences, so this chipset is introduced as an absolutely different solution with a different name, KT266A. In fact, it was nothing else but a purely marketing move: KT266A is just a CE revision of the old KT266. Its previous revision was numbered as CD. Nevertheless, the difference in marking between KT266A and KT266 North Bridges comes not only to the revision number. The North Bridge of KT266A is called VT8366A, while that of KT266 was marked as simply VT8366.

The only remakes in VIA KT266A were introduced in the memory controller. As a result, its performance improved, and the whole chipset should now work notably faster. All the other features of the North Bridge, as well as all the parts of the South Bridge remain the same. Thus, it makes no sense to describe the abilities of VIA KT266A. If you're eager to learn more about the characteristics of KT266 and KT266A, please, click here for this info.

Just let us have a closer look at what VIA improved in the DDR SDRAM controller integrated in KT266A. Now it is called "Enhanced Memory Controller with Performance Driven Design". In comparison with the previous KT266 version, the DDR SDRAM controller of KT266A has better timings. It results in lower memory latency and greater effective bandwidth. Data queues have also been deepened (this allows to process a bigger number of operations at a time and this way to decrease the latency). Then, the new controller is able to transfer up to eight quad words per clock (i.e., twice as much as the controller of KT266 could).

As for the formal specs, KT266A follows KT266 step-into-step: it supports up to 3GB PC2100/PC1600 DDR SDRAM and PC133/PC100 SDRAM (4GB in case of registered memory modules) and allows for asynchronous CPU and memory bus clocking. A similar updated DDR SDRAM controller VIA implemented in its new DDR core logic for Pentium 4 processors, P4X266.

Since KT266A is 100% pin-compatible with KT266, mainboards based on the new chipset are to start shipping shortly after VIA commences to ship KT266A in mass. It means that in less than a month we have a good chance to meet KT266A mainboards available in retail. The old KT266 will stop shipping as soon as KT266A starts. For this reason, we advise you to be careful buying a KT266 or KT266A based mainboard this autumn: don't forget to check which North Bridge microchip is really installed on it.

As we have already said, the key feature of VIA KT266A is supposed to be its comparatively better performance. Let's inquire what our tests have shown.

Test Methods

For this test session we got hold of a reference mainboard based on VIA KT266A. This is what it looks like:

It can be seen quite well that the reference mainboard from VIA is designed in MicroATX form-factor. It is equipped with only two DDR DIMM slots, two PCI, one ACR slot and one AGP 4x. Of course, for advanced users these characteristics are utterly poor, but as a testing sample this board suited us just perfectly. The design of mass-scale products based on VIA KT266A should be no worse than that of KT266 based ones, so we hope to stay satisfied with the abilities of the new KT266 based mainboards, which are to become available soon.

For a better comparison, we included all the other DDR chipsets for AMD Athlon CPUs in the race list. Alongside with VIA KT266A we tested the paragon AMD 760; the new successful B0 revision of ALi MAGiK 1; the fastest chipset of the previous tests, SiS 735; and the old VIA KT266. As mainboards representing different chipsets, alongside with the reference mainboard on VIA KT266A we selected the only certified AMD mainboard based on KT266, MSI K7T266; the fastest board of the latest generation on AMD 760, EPoX EP-8K7A; reference mainboards based on SiS 735 and the new revision of ALi MAGiK 1 from SiS and ALi.

The ultimate configurations of our testbeds looked as follows:

Tests in office and gaming applications were ran in Windows 98 SE. Professional OpenGL applications were launched in Windows 2000 Professional SP2.

Performance

First of all, let us see how things stand in the synthetic tests, which reveal the performance of the chipsets' memory subsystems. This information will help us to explain the results obtained in real applications.

Cachemem test serves to assess the bandwidth and latency of the DDR SDRAM controllers integrated into different chipsets. As you can deduce from what we got, VIA has worked well to increase the performance of its previous core logic. Now VIA KT266A is the winner in all the disciplines: like SiS 735, it ensures the lowest memory latency; in reading from the memory it overruns the former leader, ALi MAGiK 1; in writing it's faster than AMD 760. All this gives KT266A a really good chance to break ahead of the competitors in any other tests.

We can't fail to admit that now it is clear how unripe VIA KT266 chipset was. In order to launch KT266 to the market as soon as possible, VIA gave up any optimization headache. This resulted in the chipset's status of the slowest DDR Socket A solution.

In SiSoft Sandra VIA KT266A demonstrates equally outstanding results. This happens thanks to its low latency on the one hand, and to the fast data transfer on the other one.

The next thing to find out is how the results VIA KT266A showed in the synthetic tests correspond to its achievements in real applications.

SYSmark 2001 checks how fast the systems are in the most popular office and internet content creation applications. This benchmark confirms that VIA KT266A is really faster than KT266. The performance growth amounts to circa 2%, but it is quite enough for KT266A to join the leaders - the new revision of ALi MAGiK 1 and SiS 735.

Internet Content Creation part of SYSmark 2001 demonstrates that KT266A, as well as KT266, occupies the leading positions in tasks of the kind (content creation applications).

According to this part of SYSmark 2001, which measures the performance in typical office applications, KT266A is faster than KT266 by 4%, but this time it cannot catch up with the leaders, SiS 735 and ALi MAGiK 1. Surely, it's not for the memory controller alone, because in office applications the performance is determined not only by the capabilities of the CPU and memory subsystem. It is much influenced by the disc subsystem too. Apparently, this is a weak point of KT266A.

Business Winstone 2001, an analog to SYSmark 2001, indicates that KT266A is 8.5% faster than KT266 - a marvelous result that brings KT266A to the top of the list.

Content creation applications are most sensitive to the memory bandwidth. That's why the rating on this diagram corresponds to the real bandwidth we assessed, for example, in the synthetic Cachemem test. No wonder, VIA KT266A outraces all the other chipsets including the former leader, SiS 735.

Aiming to get an idea about the chipsets' performance in office applications, we measured the time they needed to compress a huge amount of data with the popular WinZIP 8.0. The idea of this test is that apart from constantly addressing the memory, the compression utility keeps on working with the disc subsystem. These test results will enable you to assess the efficiency of the buses used in different chipsets to connect the Bridges, as well as the quality of the IDE Bus Master drivers. The diagram above shows how long it took the systems to compress a directory with Unreal Tournament. The shortest time stands for the best result.

Thanks to its fast memory subsystem, VIA KT266A performs much better than the previous revision of KT266, though SiS 735 is too good for it. We believe, the performance of KT266A will improve as soon as new versions of the IDE Bus Master drivers come out.

At last, let's have a look at what KT266A is worth in games.

Results in Quake3 have always been strongly determined by the real memory bus bandwidth. This leads to heavily diversified statistics. The improved memory controller of VIA KT266A brings it the first prize. Mind the fact that VIA KT266A is good 12.5% ahead of the first DDR chipset for the Athlon CPUs, AMD 760. Who could have expected that improved timings and buffering mean so much for a core logic?

As the resolution in Quake3 grows, the results level out. But even in 1280x1024x32 mode, where the main strain is produced upon the graphics accelerator, chipsets with faster memory controllers still show higher performance.

A similar situation is observed in another game, Unreal Tournament. VIA KT266A, being 8% faster than the regular KT266, is leading in this test as well.

At higher resolutions reading from the memory becomes the key factor. That's what enables VIA KT266A to win again and to widen the performance gap, which separates it from the other racers.

In Serious Sam we can trace the same trend. As nearly always, VIA KT266A proves the fastest DDR SDRAM chipset for AMD Athlon CPUs.

Like in Quake3, at higher resolutions VIA KT266A doesn't show the best result. First of all, it must be the vxd driver to blame, but not the chipset itself.

3DMark 2001 tests give us a chance to see how the chipsets feel in upcoming games, which make use of DirectX 8. Whatever the difference in the software is, we haven't noticed any decisive changes in the results - they are very close to what we saw in the other gaming tests.

These test results generate no situation twist either.

In the end we tested our platforms in SPECViewPerf. This software emulates the work of professional OpenGL applications. They are also very sensitive to the real memory bus bandwidth, so these results are particularly interesting for us. As we expected, in tasks of the sort VIA KT266A reveals really high performance. In spite of the modest results obtained by the old KT266, KT266A runs notably faster than all its rivals.

Conclusion

The basic conclusion we can make after analyzing the test results is that VIA has finally succeeded to create a DDR Socket A core logic with acceptable (to be more exact, unprecedented) performance. Sadly, at first VIA failed to tackle this problem, so the hastily cooked up VIA KT266 is an abortive product.

Now everything is peachy:

• VIA will offer the fastest Socket A core logic for AMD Athlon CPUs with DDR SDRAM support. KT266A is to become outrageously popular among the owners of Socket A systems, at least until mainboards with official PC2700 DDR SDRAM support crop up.
• SiS 735 that impressed us during the previous test session, but now looks not so smart after our acquaintance with VIA KT266A, will be positioned as a value DDR solution. In this sector SiS 735 is likely to enjoy great popularity thanks to the low cost mainboards it allows to manufacture.
• AMD will most probably stop shipping its AMD 760 in the nearest future. This core logic has already fulfilled its mission: for almost a year it has served as an instrument for the DDR technology popularization.
• ALi will have equal chances in the value sector with SiS, but only in case it solves the problem with mass-scale shipments. If so, the new B0 revision of ALi MAGiK 1 is smart enough to become a success.

However, the battle blaring at the market of DDR chipsets for AMD Athlon CPUs is not over yet. Now we are looking forward to the arrival of NVIDIA nForce with a 128-bit memory bus and chipsets with official PC2700 DDR SDRAM support.