by Ilya Gavrichenkov
04/27/2001 | 12:00 AM
Watching Intel and AMD, that are busy promoting two new types of memory, Rambus and DDR SDRAM, we remember what people usually say: the first impressions are the most lasting. RDRAM is the only memory supported by Pentium 4 platforms, but it doesn't warm up the CPU's popularity in any way.
Mainboard manufacturers as well as ordinary users are all looking forward to Brookdale core logic due in the second half of the year. It is to tie Pentium 4 with PC133 SDRAM. Thus, RDRAM appears totally neglected and this is no technology matter. Pentium 4 is not the same as Pentium III: in Pentium 4 systems it has become reasonable to increase the memory bus bandwidth. Moreover, the distressingly big latency - a grave drawback of Rambus architecture - is not so obvious on Pentium 4 platforms thanks to the dual-channel memory access with interleaving. Besides, the price for RIMM is now not that exorbitant as it used to be a year ago. From an unbiased point of view, RDRAM is not so bad as many of us believe it to be. The problem is absolutely different. It is Rambus company itself.
In fact, Rambus scared away its potential partners bringing endless suits to memory manufacturers and persistently demanding license fees. As a result, the share of RDRAM on the market gets no higher than 10%, though Rambus has been promoting it for quite a long time already. Only these days, when Intel is making titanic efforts to popularize this memory, the standstill seems to be overcome.
DDR SDRAM makes a perfect contrast to Rambus. Since it is an open standard developed by several companies, DDR SDRAM enjoyed vigorous advertising long before the first DDR platforms came into being. Among the promoters there were many well-known chipset and memory manufacturers including AMD and many others.
This way, they created a strongly positive attitude towards the new product, although it had its lows too. This hearty welcome wasn't spoilt even with really too high price that didn't correspond to the modest performance gain provided by this memory type. Everyone was waiting for some supreme force to interfere and kept on waiting. Meanwhile the memory manufacturers pledged to cut the prices for DDR SRDAM when there appear enough mainboards supporting this type of memory. In their turn, the mainboard manufacturers were waiting for more chipsets to come out, while the chipset manufacturers lingered for some reason. In fact, it all mostly depended on only one chipsets manufacturer, VIA, which had a nearly unique opportunity to produce core logic in mass. However, the company's DDR pioneer, VIA Apollo Pro266, wasn't enough to settle all this fuss, as long as the architecture of Pentium III, for which this core logic was intended, didn't allow increasing the performance by raising the memory bus bandwidth. So, VIA Apollo Pro266 could hardly appeal to the public. A different thing was VIA KT266 meant for AMD Athlon CPUs. Presumably, Athlon systems with DDR SDRAM support could provide a notable gain in performance. But the already released DDR chipsets for this platform - ALi MAGiK 1 and AMD-760 - brought no significant performance breakthrough. Nonetheless, VIA KT266 was still an apple in the hardware world's eye and finally the first mainboards based on it appeared. Well, we couldn't ignore an event like that and now you can scroll through the most detailed analysis of this chipset's performance crowned with our severe but fair judgment.
VIA has already announced and is now shipping one chipset for AMD Athlon processors with 266MHz bus. We mean KT133A, a slight recast of VIA KT133. At the same time, KT133A doesn't support DDR SDRAM, that is why VIA made up its mind to release a second chipset for Athlon CPUs with 266MHz bus, KT266, supporting the new memory type. To do so it was quite enough to redesign the memory controller of KT133A, but VIA chose another method. KT266 represents a new generation chipset, differing dramatically from all the previous chipsets for AMD processors. Just take a look at the scheme for VIA KT266 and you'll understand what we mean:
It requires no academic education to see that VIA KT266 is architecturally closer to VIA Apollo Pro266 Pentium III chipset than to VIA KT133A. The reason is the second major innovation implemented in KT266, which we have already met in VIA Apollo Pro266: there is a special V-Link bus connecting the Bridges instead of the formerly used PCI bus. The bandwidth of V-Link is two times greater than that of PCI totaling 266MB per second. This freed the chipset from its potential bottleneck, caused by the increasing demands of the South Bridge controllers. In particular, the integrated two-channel ATA/100 controller, a USB controller supporting up to six USB ports, AC'97 and 10/100 BaseT Ethernet controllers could've easily overload the old PCI bus if they transferred the data simultaneously.
The introduction of V-Link caused some other changes in the chipset structure. The PCI controller, previously located in the North Bridge, moved over to the South Bridge, for its place in the North Bridge appeared occupied by the V-Link controller. Subsequently, the architecture of VIA KT266 now resembles that of Intel's chipsets, where the North Bridge supports only the CPU bus, the memory and AGP 4x.
The memory controller built in the North Bridge of KT266 looks like that of VIA Apollo Pro266. It supports almost all the currently available types of SDRAM including PC2100/PC1600 DDR SDRAM and PC133/PC100 SDRAM. That's why good many mainboard manufacturers are about to create products based on VIA KT266, equipped both with 184-pin slots for DDR SDRAM and 168-pin slots for the common SDRAM. Due to this combination, the on-going solutions will grant certain flexibility in choosing the memory type and make the upgrade of PC133 SDRAM platforms much easier.
One of the most important features of VIA KT266's memory controller is its ability to work asynchronously with the CPU. Accordingly, you may take a 133MHz memory (PC2100 DDR SDRAM or PC133 SDRAM) and a CPU clocked at 100MHz FSB. Apart from that, VIA KT266 allows using PC1600 DDR SDRAM in systems with 133MHz FSB.
VIA KT266 is the only DDR chipset for Athlon CPUs bundled with an asynchronous memory controller. Neither AMD-760, nor ALi MAGiK 1 let clock the CPU and the memory at different frequencies. On the one hand, it is good, since the users have a broader choice of possible configurations, but on the other one, there are some drawbacks.
The main problem about an asynchronous memory controller is its lower performance. An asynchronous chipset has to make sure that the data transfers along the CPU-to-memory bus are synchronized, hence instructions and transferred data are to be pipelined and interlaced with additional wait states. Because of all this, an asynchronous memory controller cannot theoretically be as fast as an "ideal" synchronous one.
In the case discussed we are mostly concerned about the performance of KT266 compared with another chipset for Athlon, AMD-760, which does feature a synchronous memory controller. It is still too early to assume that the asynchronous KT266 is slower than AMD-760 with its synchronous memory controller. In KT266 VIA has implemented an algorithm, which should compensate for the upsettingly high latency of its asynchronous controller. What we mean is the memory interleaving. The idea is to address the memory banks in an interleaving manner, thus sometimes starting to transfer a new portion of data before the previous one is fully transferred. There is nothing of the kind in AMD-760. Instead, it boasts super bypass - a pattern, which helps to lower the memory controller latency by rejecting some operations with PCI and AGP buses when the memory is engaged. As you see, it's hard to say at once, which of the two chipsets - AMD-760 or ALi MAGiK 1 - works faster with the memory.
Nevertheless, even if we prove that VIA KT266 is much slower than AMD-760, it won't make us dub it a failure. There are no doubts about VIA KT266's popularity: just look how enthusiastically the mainboard manufacturers are announcing their new solutions based on this core logic. To our great disappointement AMD-760 suffers two serious shortcomings preventing it from becoming widely spread.
First, AMD doesn't strive to become a chipset manufacturer. AMD-760 is its pilot product which is supposed to simply demonstrate the capabilities of DDR SDRAM and new Athlon CPUs with 266MHz bus. For this reason AMD is unable and unwilling to release this chipset in large quantities, enough to satisfy the demand of mainboard manufacturers. AMD wants to see other companies as market providers, keeping in mind first of all VIA and ALi. And as far as ALi yet has no opportunity to ship its ALi MAGiK 1 in demanded quantities, the only real manufacturer of DDR chipsets for AMD processors remains VIA offering its KT266.
The second reason for VIA KT266's undoubted success are the prices for products based on this core logic. Mainboards built on VIA KT266 are cheaper than those based on AMD-760. It can be explained by the lower price for KT266 chipset: $32 versus $40 for AMD-760. Besides, AMD-760 requires 6-layer mainboard design, which results into a more complicated PCB. VIA KT266 in its turn lets the manufacturers get away with the traditional four-layer design. As a result, mainboards based on KT266 are awaited to be pricing $20 cheaper than their counterparts based on AMD-760.
At last, let us compare the formal characteristics of KT266 and its main competitors supporting AMD Athlon CPUs with 266MHz bus:
(AMD-761 + VIA 686B)
|VIA KT266 |
(VIA VT8366 + VIA VT8233)
|ALi MAGiK 1 |
(ALi M1647 + ALi M1535D+)
(VIA VT8363A + VIA 686B)
|Processor bus||266/200MHz EV6||266/200MHz EV6||266/200MHz EV6||266/200MHz EV6|
|Memory||PC2100/PC1600 DDR SDRAM||PC2100/PC1600 DDR SDRAM |
|PC2100/PC1600 DDR SDRAM |
|Max memory size||2GB (4GB in case of Registered DIMM)||3GB (4GB in case of Registered DIMM)||3GB||1.5GB|
|Asynchronous memory bus||-||+||-||+|
|Max number of PCI Master||7||5||6||5|
|Bus between the chipset bridges||PCI (133MB/sec)||V-Link (266MB/sec)||PCI (133MB/sec)||PCI (133MB/sec)|
For ample information about VIA KT266 we addressed those manufacturers who are now ready to present their mainboards based on this core logic. As we figured out, most of these products are far from being on the verge of mass production and are scheduled for no sooner than May. However, three companies (EPoX, Gigabyte and Shuttle) have managed to send us their samples to test. So, these three ones are most likely to be in the vanguard of KT266 mainboards. As for MSI, which mainboard built on KT266 is already on sale in some places, we received it a bit later and will offer you a separate review in the nearest future.
Let us take a keen look at the participants of our investigation:
|EPoX EP-8KHA||Gigabyte GA-7VTX||Shuttle AK32|
|Supported CPUs||AMD Athlon/Duron (200/266MHz)|
|Overclocking Friendly Features||Supports CPU Clock Multiplier Setting, Vio, Vagp, Vdimm||Supports CPU Clock Multiplier Setting, Vio|
|Memory||3 184-pin DIMM slots for PC1600/PC2100 DDR SDRAM||3 184-pin DIMM slots for PC1600/PC2100 DDR SDRAM||2 184-pin DIMM slots for PC1600/PC2100 DDR SDRAM|
2 168-pin DIMM slots for PC100/PC133 SDRAM
|AGP Slot||AGP 4x||AGP Pro||AGP 4x|
|Expansion Slots (PCI/ISA/AMR)||6/0/0||5/0/1||6/0/1|
|Integrated Sound||AC'97||Creative CT5880||AC'97|
|Additional Features||Integrated P80P Debug (POST) controller||DualBIOS technology||Integrated 10/100 Fast Ethernet controller|
|BIOS||Award Modular BIOS v6.00PG||AMI BIOS 1.24a||Award Modular BIOS v6.00PG|
|Form-Factor||ATX, 305x245mm||ATX, 305x244mm||ATX, 305x244mm|
EPoX was one of the first to create a mainboard based on VIA KT133A. Just the same can be said about KT266 mainboards. While the products of other manufacturers are for the most part being developed currently, EPoX brandishes a ready-to-serve solution, 8KHA.
Speaking of features, 8KHA is provided with all the required basic characteristics, which are to be present on a modern DDR Socket A mainboard. 8KHA has one AGP 4x slot, six PCI and three DDR DIMM slots. You'll find no outdated ISA and system-integrators-oriented ACR/AMR slots.
EPoX decided not to bundle this KT266 mainboard with IDE RAID controller, which used to be a common feature of all its previous products. On the other hand, 8KHA is equipped with a traditional Debug controller, which is specially designed to detect and brush away hardware troubles revealed in POST process.
This mainboard supports 6 USB ports and AC'97 sound controller. To implement AC'97 sound EPoX has resorted to a dual-channel codec by VIA.
Sadly, although the board has an "easy" layout of seven expansion slots as a whole, we can't but criticize some inconveniences in the design. Namely, the AGP slot is located too close to the DIMM slots, so graphics cards installed into the AGP slot may block the clips of DIMM slots. Then, though AGP has a clip to fix the graphics card in the slot, this clip's design doesn't allow installing AGP Pro cards, which can work in regular AGP slots. Moreover, IDE connectors are moved to the left, therefore IDE cables are allocated in far not the best place. The same thing happens with the ATX power supply connector. Its placing behind Socket A has already provoked heated criticism, so there is no need to add anything on this matter. But the most unpleasant point is that the capacitors of the CPU power supply circuit are put too close to Socket A. This may cause certain problems with installing large coolers, say Chrome Orb.
8KHA's BIOS Setup offers a complete set of memory optimizing settings, including manual CAS Latency and Memory Interleaving adjustment. By means of BIOS Setup one can also do some overclocking. The sample we tested allowed adjusting the FSB frequency only below 166MHz with an increment of 1MHz, but EPoX promises to improve the following versions of the BIOS and to introduce possibilities for changing the CPU multiplier, Vcore and Vio.
Like other mainboards by Gigabyte, GA-7VTX has a blue PCB. The board has three 184-pin slots for DDR SDRAM, one AGP Pro and one AMR slot. Of course, on the one hand only five PCI slots on the board seem to be not enough for today's needs, as most other manufacturers equip their products with six slots and it is not a problem to find some space for one more slot on the board. However, VIA KT266 can support no more than five PCI Master devices, so it makes no sense to implement a sixth slot - it will be invalid anyway. Thanks to an additional AGP Pro slot together with a 12V power supply connector, GA-7VTX fits for professional power-consuming graphics cards.
As well as other Gigabyte's leading products, GA-7VTX supports DualBIOS technology, which implies keeping a backup version of the mainboard BIOS on a similar Flash ROM chip.
Besides that, we wish to highlight an integrated fully-functional Creative CT5880 sound controller installed on GA-7VTX instead of the software AC'97 codec. Then, the board's chip supports the whole bunch of USB ports, i.e. the board allows six ports to be connected.
The design is pretty well-planned but for a single slip: the ATX power supply connector lies between Socket A and the ports of the case rear panel. For the sake of an optimal layout Gigabyte's engineers had to make PCB rather large in size, but there should arise no problems with installing GA-7VTX into standard ATX cases.
We would like to stress that we got hold of just a sample, but its stability turned out splendid. Actually, high stability is part of Gigabyte's image.
Lately Gigabyte decided to take care of overclockers' interests and began to introduce corresponding opportunities on its mainboards. The new GA-7VTX stands in the row. Its appreciable overclocking abilities are implemented in the form of dip-switches and jumpers. However, the board should appeal even to experienced overclockers: it allows you to change the CPU multiplier, the FSB frequency, Vio, Vagp and Vdimm. Furthermore, you can also adjust the FSB frequency via BIOS - in this case GA-7VTX gives a chance to set it with an increment of 1MHz up to 161MHz.
Unlike other mainboards we have tested, GA-7VTX works with a BIOS from AMI, not from Award. Whatever, the Setup of GA-7VTX has options for adjusting CAS Latency and Memory Interleaving too.
No doubt, Shuttle AK32 is the most interesting product in this review. The main feature that makes it outstanding is the support of DDR SRDAM and common SDRAM simultaneously. For this purpose AK32 is bundled with two 184-pin slots for PC2100/PC1600 DDR SDRAM and two 168-pin slots for PC133/PC100 SDRAM. In fact, there is nothing strange about a combination like that. KT266 supports both memory types, but the mainboard manufacturers mostly prefer not to create products of the sort, anticipating a forthcoming price cut for DDR SDRAM. Now that it hasn't happened yet, Shuttle's solution looks wise. And all the more so since by upgrading one may need a DDR mainboard support PC133 SDRAM.
Except for this innovation, AK32 has no more extraordinary features. The board is equipped with one AGP 4x slot, six PCI and one CNR slot. It may be worth mentioning that the layout includes VIA KT266 10/100 Fast Ethernet controller integrated into the South Bridge. That's why among the connectors on the rear panel of the case you'll find RJ45 connector located above the USB ports.
Other features are AC'97 sound controller and the ability to support six USB ports. Keep in mind that the bracket with two more USB ports, which is to be installed in the case rear panel is shipped together with the board in one package.
The design of AK32 can't be taken for perfect. Unfortunately, this mainboard isn't free from traditional drawbacks: the graphics card installed into the AGP slot blocks DIMM clips, while FDD and IDE connectors are moved too far away from the right edge of the board, so that the cables are hanging all through the case. Still, the ATX power supply connector is situated most favorably - on the right edge of the mainboard. Another pleasant thing is that Shuttle's engineers have left enough space around Socket A, making the use of massive coolers possible. So, there should be no troubles with good CPU cooling on AK32.
To say more, AK32 has proved to be the fastest of all the presently available mainboards. This fact joined with flexibility and traditionally low prices for Shuttle's products lets us call AK32 a great success.
Overclockers will like this mainboard too. Having no jumpers for CPU configuring, the board allows adjusting the CPU multiplier, the FSB frequency (with an increment of 1MHz up to 166MHz) and Vcore via BIOS Setup. In the same manner, you may adjust CAS Latency and enable Memory Interleaving for both memory types.
While KT266 mainboards are nothing more than preproduction samples, it doesn't make much sense to compare them with one another. By the time they appear in retail, their performance may be changed by the new version of the BIOS or some design improvements. This way, in this study we tried to make out the top performance one can sweat from systems with VIA KT266 chipset. For a better comparison, all the results were obtained on today's fastest mainboard in our lab, Shuttle AK32. Another bonus we got from using this board in our test platform was the opportunity to see the working speed of VIA KT266 with DDR SDRAM alongside with estimating the chipset's performance with PC133 SDRAM.
We compared the performance of VIA KT266 with that of other chipsets supporting AMD Athlon CPUs with 266MHz bus, AMD-760, ALi MAGiK 1 and VIA KT133A. Since ALi MAGiK 1, as well as VIA KT266, supports both DDR SDRAM and common SDRAM, we tested this core logic in both configurations.
Eventually, we assembled six test systems:
PC2100 DDR SDRAM
|ALi MAGiK 1 |
PC2100 DDR SDRAM
|ALi MAGiK 1 |
|VIA KT133A |
|VIA KT266 |
PC2100 DDR SDRAM
|CPU||AMD Athlon 1.2GHz (266MHz FSB)|
|Mainboard||Gigabyte GA-7DX||ASUS A7A266||ABIT KT7A||Shuttle AK32|
|Memory||256MB PC2100 DDR SDRAM||256MB PC133 SDRAM||256MB PC2100 DDR SDRAM||256MB PC133 SDRAM|
|Graphics Card||Creative 3D Blaster Annihilator 2 Ultra (NVIDIA GeForce2 Ultra)|
|HDD||IBM DTLA 307015|
The gaming and office tests were run in Windows 98 SE and professional OpenGL tests - in Windows 2000 Professional SP1.
Before passing over to real applications, we resorted to SiSoft Sandra 2001 to check the performance of the memory subsystem.
Although very often SiSoft Sandra 2001 benchmark is accused of the results instability, we wouldn't trim it so fiercely. First, tests of the memory bus bandwidth, which are made by SiSoft Sandra 2001, are based on the conventional STREAM algorithm. Second, if there is any instability in the results, it is caused by the operating system (we mean Windows). Thus, you had better make a restart before launching Sandra 2001 to clear the swap file. In this case you are sure to have adequate performance.
Discussing the outcome, we can note right away that VIA KT266 hasn't proven up to our expectations. It shows the lowest results both with DDR memory and common SDRAM. Strikingly enough, the performance of VIA KT266 working with PC2100 DDR SDRAM turns out lower than that of VIA KT133A supporting only PC133 SDRAM.
The second tour of Sandra 2001 shows the memory bus bandwidth when the FPU carries out some operations with it. As long as FPU operates with longer data structures, these results depend on the memory bus bandwidth much more than in the previous case. No wonder that here all chipsets with DDR SDRAM outpace those with PC133 SDRAM. But VIA KT266 again has not much to be proud of: it is slower than all its rivals working with the similar memory type.
Mind the fact that the distressingly low results of VIA KT266 don't at all mean that the chipset will perform poorly in real applications as well. The results of Sandra 2001mostly depend on the peak memory bus bandwidth, but not on the memory latency. In the majority of real tasks the performance is equally determined by both these factors.
This test illustrates our latest statement. Business Winstone 2001 checks how fast systems can work in typical business applications. This kind of tasks involves intense work with small amounts of data. Therefore, memory subsystem latency should matter a lot now. And here VIA KT266 has no problems. The only rival, which managed to surpass VIA KT266, appeared VIA KT133A combined with PC133 SDRAM.
It is quite natural, in fact, because the latency of PC2100 DDR SDRAM is generally greater than that of PC133 SDRAM (we have already mentioned it here). This factor explain the results we obtained in Business Winstone 2001:
Content Creation Winstone 2001 follows the same working principle as the previous test, but the results are absolutely different. There is nothing to be surprised at: these application are designed for content creation, that is for working with large amounts of data. Correspondingly, memory bus bandwidth comes in the forefront. We guess that no one will be confused to see the results in Content Creation Winstone 2001resemble those in SiSoft Sandra 2001: VIA KT266 lags behind both its DDR rivals, AMD-760 and ALi MAGiK 1.
By the way, VIA KT133A with PC133 SDRAM keeps on showing really good results.
The basic difference between this test and the previous two ones is that this benchmark checks the performance in successively, not simultaneously executed applications. Unlike the Winstone 2001 tests, it loads the memory bus less tangibly and the performance is now influenced by other factors: the latency, the speed of IDE controller and so on. That's the ground for the following results: VIA KT133A is the leader, then there come VIA KT266 and ALi MAGiK 1. The performance of AMD-760 is the poorest this time.
We couldn't get away without the classical archiving testing in WinZIP. To check the archiving speed, we measured the time needed for zipping a directory with Unreal Tournament. As you've guessed, shorter time stands for better results.
The performance of our test systems in WinZIP is of interest to us due to two factors. Firstly, it depends on the memory subsystem performance. Secondly, it's the speed of the IDE controller that counts as well. The leading positions are eventually occupied by VIA chipsets and AMD-760 (by the by, our AMD-760 based mainboard featured VIA 686B chip as a South Bridge), and ALi MAGiK 1 is an apparent outsider.
In 3DMark2001 VIA KT266 keeps abreast with ALi MAGiK 1 and falls slightly behind AMD-760. This benchmark indicates a pure victory of DDR chipsets over their SDR competitors, thus confirming the fact that for future 3D games, which are modeled in this test, high-performance memory bus will be required.
To test the chipsets in Quake3 we chose an updated version (Quake3 1.27g) and a new gaming demo127, which loads the system harder than the formerly used demo001.
We have already mentioned that at low resolutions Quake3 is sensible to the memory latency and bus bandwidth. That is why our bosom-readers mustn't be astonished to see the results we've got: AMD-760 makes the most of its well-optimized memory controller to break ahead of KT133A with PC133 SDRAM, which contrives to provide the shortest access time, and VIA KT266 is the third in the list.
Higher resolution brings no decisive twist. It means that the AGP bus, which is a crucial item in this benchmark, is equally smart at all the chipsets. None of the tested samples has any problems with the AGP 4x bus bandwidth.
Once again we observe the same state of things. VIA KT266 is overtaken by both AMD-760 and VIA KT133A. As long as KT133A works with PC133 SDRAM and its memory bus bandwidth can't grow further, we can explain the success of this chipset with lower latency, which is also the matter for AMD-760.
Higher resolution doesn't swing the general trend.
In our tests we have included a new game called Serious Sam built on the latest OpenGL engine created by Croteam. In spite of this innovation, the outcome is much like that of the previously used games. VIA KT266 keeps up with ALi MAGiK 1, while VIA KT133A and AMD-760 are again faster than the new DDR solution from VIA.
As the resolution grows, VIA KT266 gets a little bit ahead of ALi MAGiK 1. All the other figures are more or less leveled out, signaling limited performance of the graphics systems. But as a whole, the situation remains unchanged.
Mercedes-Benz Track Racing is a fresh game by DirectX, using a great number of textures. Therefore the fps rate depends on the memory subsystem performance quite a lot. Surprising as it might seem, VIA KT266 lags behind all the racers.
The same thing can be said about the results obtained at higher resolutions.
Professional OpenGL applications are good to evaluate the memory bus bandwidth. Here all DDR chipsets outperform SDR solutions. However, we were very upset to see VIA KT266 be the slowest DDR chipset in SPECviewperf.
The tests have brought no optimistic evidence for the performance of VIA KT266. This so much cherished DDR chipset for Socket A platforms has turned out close to the slowest ALi MAGiK 1, which is sometimes surpassed by VIA KT133A with PC133 SDRAM. However, before putting a gravestone over VIA's child, let us try to figure out the reasons for its deplorable performance.
Obviously, the key problem is the memory subsystem performance. In order to prove it, we made use of Cachemem utility, which can help to measure the memory peak bandwidth by reading and writing as well as its latency (i.e., how long it takes the memory to respond to a request). The results are plotted in the following schemes:
It's easy to notice that although the memory latency of VIA KT266 with DDR SDRAM is quite all right, being slightly higher than the latency of AMD-760, it leaves much to be desired when it comes to transferring the data. KT266 can write data to the memory only a trifle faster than ALi MAGiK 1, and in reading KT266 is a real failure, lagging behind all the other chipsets by more than 30%. Judging by nearly the same reading speed in case of both: DDR SRDAM and PC133 SDRAM, we can assume that this platform has some artificial hindrance preventing it from showing proper results.
KT266's future now totally depends on the nature of this hindrance: it may be for the chipset itself (then one can do nothing about it) or for some particular mainboard or its BIOS. In an attempt to find out the truth, we tested some other mainboards built on VIA KT266 by Cachemem benchmark. Frankly speaking, the results were more than shocking. In particular, Gigabyte GA-7VTX showed normal read speed (circa 1000MB/sec), but the memory latency was even greater than that of ALi MAGiK 1 based mainboard.
This argues for the "innocence" of the chipset. If the engineers manage to get the same read speed (from the memory) as we saw by Gigabyte GA-7VTX, having preserved the outstandingly low latency, then mainboards based on VIA KT266 will be likely to have much better performance than the samples we had at our disposal.
At the moment we can do nothing but wait. Let' hope that soon some improved products based on VIA KT266 will appear. In our turn, we promise to write an update to this article as soon as we get hold of such a product.
Well, at present VIA KT266 core logic still needs a thorough remake, though even now it's clear that it will be a success one day. Ultimately, mainboards built on KT266 are expected to catch up with the today's #1, AMD-760, in performance, and their lower price will attract the buyers. So, it looks very much as if VIA might be able to strengthen its positions as the leading Socket A chipsets manufacturer.
For the time being, PC2100 DDR SDRAM remains much more expensive than PC133 SDRAM. It will surely tell on the popularity of DDR chipsets in general and KT266 in particular, setting them back from the heights of VIA KT133A. Especially, since VIA KT133A proves not that radically slower than DDR platforms.
Meanwhile, those who wish to have a high-performance system based on Athlon CPUs, should better decide on AMD-760 based boards. And if you are more concerned about the price-to-performance ratio, VIA KT133A will be the best choice.
To be continued...