Anticlimactic would be an appropriate word to describe the launch of the very first Socket-A chipset with support for DDR SDRAM, the AMD 760.  The AMD 760 was released almost six months ago, and unfortunately its introduction was not met with much enthusiasm among those in the community simply because the price premium the chipset and its new memory commanded was not worth the 0 - 10% performance improvement it offered over present day PC133 solutions.

However, AMD's motives behind releasing the 760 chipset were not to show the immediate benefits of DDR SDRAM nor were they to sell hundreds of thousands of 760 chipsets.  Instead AMD's goal, as a CPU manufacturer and not a chipset manufacturer, was to promote the technology that their CPUs would need in the future.  By being a driving force behind DDR SDRAM technology, AMD took a strong position in encouraging the rest of the market, particularly 3rd party chipset manufacturers such as ALi, SiS and VIA, to adopt that technology in their upcoming solutions.

While we all knew a solution from VIA was in the works, ALi was actually the first to offer a DDR SDRAM chipset for the Athlon platform after the AMD 760 was released.  Unfortunately, the ALi MAGiK1 didn't impress us when we were first introduced to it last November, and in our most recent comparison with a much more mature platform, the chipset has yet to shine as even an alternative to the current PC133 chipsets.

Speaking of which, it was a PC133 Socket-A chipset that took off in popularity instead of the AMD 760 at the end of last year.  As you all probably know, that was the KT133A and it is still quite a popular solution simply because it offers compatibility with PC133 SDRAM and provides very low latency operation as we noticed in our latest comparison.  Its low latency operation allows it to excel quite well in today's applications and games which is represented accurately in most of the benchmarks.

VIA does understand that they cannot live off of the laurels of their PC133 platforms forever so it came with no surprise that in the first month of the new year they announced volume shipments of their KT266 chipset, VIA's first DDR solution for the Socket-A platform. 

What's wrong with the 760?

As we found out in our latest Socket-A Chipset Comparison, the AMD 760 chipset is actually a very impressive chipset.  AMD has come a long way since the days of the AMD 640 (which was actually the VIA VP2 in disguise) to the point where they are a very capable chipset designer even though they still refuse to take the next step to become a manufacturer as well. 

This brings up the first problem with the AMD 760 chipset: it isn't a high volume product.  This unfortunately means that it is a more expensive solution to produce, and it also means that motherboard manufacturers are less likely to adopt and promote designs based on the chipset due to its relatively limited production nature.

In mentioning the first problem with the AMD 760 chipset, we actually introduce the second issue with the AMD 760 chipset, which is the problem of cost.  AMD 760 based motherboards already require the use of 6-layer PCBs (the more layers, the greater the cost) as opposed to the 4-layer PCBs manufacturers were able to get away with using the KT133/A chipsets.  This fact alone scared away quite a few motherboard manufacturers, such as ABIT, that find it very hard to swallow the added cost of producing a 6-layer board in an already very tight profit margin market. 

The PCB requirements combined with the added cost of the termination circuitry and AMD 760 chipset itself made the solution far less attractive to motherboard manufacturers.  It is because of the cost issue that many motherboard manufacturers used the VIA 686B South Bridge instead of the AMD 766 South Bridge which is slightly more expensive. 

Luckily, from a performance and feature set standpoint there aren't any problems with the AMD 760 chipset.  If you'll remember the 760's predecessor, the AMD 750 chipset, not only did it not offer support for PC133 SDRAM but it also had numerous AGP performance issues under Windows NT.  The 750 did not support AGP 4X, however we all know the usefulness of that over AGP 2X in real world situations.

So while the 760 doesn't have the same shortcomings as its predecessor in terms of performance and feature-set, it does share the same Achilles heel in terms of cost which is what allows solutions like the KT266 to be so attractive.

The Chipset

While the KT266 is VIA's first DDR Athlon chipset, the Apollo Pro266 was actually VIA's first DDR capable chipset, but it was for the Socket-370 platform thus supporting the Intel Pentium III and Celeron processors.  The Pro266 was a significant departure from previous VIA chipsets because it introduced a new bus connecting the North and South Bridges in the chipset. 

As a quick refresher, the North Bridge is what houses the memory and AGP controllers. It is also the chip that lies connects directly to the CPU, thus controlling your FSB frequency among other things related to your CPU bus interface.

The South Bridge handles any ISA bridges, the integrated hard drive controller, it contains the USB controller and in the case of VIA's 686A, it also houses the keyboard/mouse controllers.

Prior to the Pro266, the North and South Bridges were connected using the 32-bit PCI bus running at 33MHz for a peak theoretical bandwidth of 133MB/s.  With current IDE hard drives approaching 40MB/s sustained transfer rates, USB devices becoming increasingly more popular and 100Mbit Ethernet cards eating up this 133MB/s of available bandwidth there is a need to remove this potential bottleneck from the system.

VIA removed this bottleneck through the use of their 8-bit V-Link bus that operates at 266MB/s, identical to the bandwidth offered by Intel's interlink bus.  V-Link takes the place of the PCI bus between the North and South Bridges, leaving the PCI bus as an extension off of the South Bridge instead of the path for communication between the two major parts of the chipset. 

You can easily see the V-Link connection between the North and the South Bridges on the Pro266 Reference board pictured below.

Click to Enlarge

The Apollo Pro266 was the first VIA chipset to feature V-Link, and the KT266 is the second.

Click to Enlarge

In order to support DDR SDRAM and V-Link, the KT266 must use a different North Bridge from the KT133/A series of chipsets however it can borrow the same 8233 South Bridge from the Apollo Pro266.  The KT266 uses the 8366 North Bridge, which offers the same feature-set as the KT133A with the exception of its DDR SDRAM support.  This does mean that the North Bridge supports PC133 SDRAM, a feature that you may see some OEMs take advantage of to provide "one-size fits all" motherboard designs.

Since the chipset uses the same 8233 South Bridge as the Pro266, we can quote our Pro266 review in regards to the features it supports:

The South Bridge is virtually identical to the 686B from VIA with the exception that it now contains the PCI bus interface, an integrated 10/100 Ethernet controller and supports up to 6 USB ports.

One of the most attractive selling points the KT266 chipset does have is that it is around 10 - 15% cheaper than the AMD 760.  While that doesn't translate into much savings for the consumer, it does help save the motherboard manufacturer quite a bit, making it a very attractive solution for them. 

The Board

One of the major reasons we conducted our latest Socket-A Chipset Comparison was because we had three different Socket-A chipsets on three different motherboards, all manufactured by the same company, ASUS.  This helped eliminate a variable that is almost always present in such comparisons, because various manufacturers place different values on optimizing their designs and BIOSes for performance or stability, etc. 

Unfortunately, ASUS has yet to produce a KT266 based motherboard and it is unlikely that they will currently because of the fact that they are pushing their A7A266, an ALi MAGiK1 based solution, as their AMD 760 alternative.  Our only option was to use the MSI K7T266 Pro that isn't the best comparison we could make to three ASUS motherboards, but it's the best we can do with what we're currently dealing with.

However as we've seen in the past, MSI boards aren't too far away from the performance of their ASUS counterparts so you shouldn't expect to see too large of a performance gap simply because of board manufacturers. 

Click to Enlarge

The K7T266 Pro is an interesting board, we weren't too happy with the BIOS setup as MSI decided to use an AMIBIOS utility for the board instead of their usual AWARD setup.  The AMI utility had some interesting quirks such as placing the clock multiplier and FSB adjustment functions underneath the Hardware Monitoring menu in the BIOS, but at least the board let's you adjust the clock multiplier unlike both of ASUS' DDR solutions.

Another BIOS complaint was that the current versions of the AMI DOS flash utility don't support the KT266 chipset and won't let you flash the BIOS of the board.  That would normally be ok because MSI does have an on-line BIOS update utility known as Live BIOS, unfortunately the link to the updated BIOS in MSI's Live BIOS script was broken at the time of publication so we couldn't upgrade our first K7T266 Pro to the shipping BIOS revision.  Luckily our friends over at NewEgg were able to send us a retail board that we used for our tests.

Other issues involved the system locking up under Windows 2000 when running SPECviewperf every now and then, as well as failure to POST after a warm reboot.  The only solution to the warm reboot problem was to unplug the test bed and then cycle power that way.  This seems to be a problem with a handful of MSI boards; although this is the first time we have been able to duplicate it in lab.  Apparently quite a few of our readers have expressed complaints with issues that resemble this warm reboot/POST failure issue.  We'll be in touch with MSI and we will work with them on a solution since we are finally able to duplicate it in the lab.

The Test

Windows 98SE / 2000 Test System
Hardware
CPU(s)	AMD Athlon "Thunderbird" 1.0GHz AMD Athlon-C "Thunderbird" 1.0GHz
Motherboard(s)	ASUS A7V133 ASUS A7M266 ASUS A7A266 MSI K7T266 Pro
Memory	256MB PC133 Corsair SDRAM (Micron -7E CAS2) 256MB Corsair PC2100 DDR SDRAM
Hard Drive	IBM Deskstar 30GB 75GXP 7200 RPM Ultra ATA/100
CDROM	Phillips 48X
Video Card(s)	NVIDIA GeForce2 Ultra 64MB DDR (default clock - 250/230 DDR)
Ethernet	Linksys LNE100TX 100Mbit PCI Ethernet Adapter
Software
Operating System	Windows 98 SE Windows 2000 Professional SP1
Video Drivers	NVIDIA Detonator3 v6.50 @ 1024 x 768 x 16 @ 75Hz NVIDIA Detonator3 v6.50 @ 1280 x 1024 x 32 (SPECviewperf) @ 75Hz VIA 4-in-1 4.29V was used for all VIA based boards (w/ 580_BETA IDE Driver)
Benchmarking Applications
Gaming	Unreal Tournament 4.32 Reverend's Thunder.dem Quake III Arena v1.27g demo127.dm3 Expendable timedemo Mercedes-Benz Truck Racing Timedemo Serious Sam Public Test 2 (Coop Party 04)
Productivity	BAPCo SYSMark 2000 Benchmark Studio Beta 2.0
Professional	SPECviewperf 6.1.2
Synthetic	SiSoft Sandra 2001 Linpack Cachemem

Memory Bandwidth - Linpack

Due to the sheer number of chipsets we're comparing in this article we didn't feel it would be too helpful to provide the usual line graph diagramming performance under Linpack as the size of the Linpack dataset increases beyond that of the Athlon's caches.

Instead, we are providing the below chart of floating point performance using a 2MB matrix size, indicating that the performance here is limited mostly by memory bandwidth and latency.

The KT266 doesn't start out too strong, in fact it actually falls behind the KT133A and the MAGiK1 with PC133 SDRAM.  We would normally expect it to perform close to the two other DDR solutions on the chart, however this is far from an indication of the overall performance picture.

Memory Bandwidth - SiSoft Sandra 2001

Sandra 2001 uses the STREAM benchmark core for its memory bandwidth tests, arguably one of the only truly useful benchmarks from the suite.  The reasons the Sandra memory bandwidth results are so useful is because they are a combination of raw memory bandwidth performance and real world usage patterns since there is a significant amount of data manipulation occurring that results in the scores produced, instead of a raw read and write performance measurement. 

Since it is rare that you only read or write data without any sort of manipulation of it, this is a more realistic benchmark, although it is still a synthetic performance measurement utility.

Again, we see the KT266 with PC2100 DDR SDRAM, coming in below the MAGiK1 and AMD 760 and in this case, even the KT133A in terms of memory bandwidth.  The difference between the KT133A and KT266 isn't great, however you would expect it to be much larger and in favor of the KT266, not the other way around.

One thing that you do have to take into consideration is that the KT133A chipset is closely based on the KT133 core that has been around for close to a year now, while the KT266 is using a brand new memory controller from VIA. 

The FPU STREAM test from Sandra uses a dataset that is twice the size (FPU calculations are generally more memory bandwidth intensive b/c of larger datasets), thus increasing the dependency of the test on memory bandwidth performance resulting in higher figures in the above chart. 

Because of the larger focus on memory bandwidth, even the matured KT133A chipset must fall back and make room for the KT266 with PC2100 DDR SDRAM however it still offers less bandwidth than the AMD 760 and ALi MAGiK1 solutions.

Memory Bandwidth - Cachemem

Cachemem is the exact opposite when it comes to synthetic memory performance tests in that the benchmark measures the peak read and write bandwidth figures without heavy data manipulation. 

Here we see that the peak performance of the KT266 running at 1000/133 with PC2100 DDR SDRAM is very close to that of the AMD 760 and the ALi MAGiK1.  It is also interesting to note that running the KT266 at 1000/133 with PC1600 DDR SDRAM vs. 1000/100 with PC2100 DDR SDRAM produces nearly identical bandwidth numbers.  Depending on whether or not the two configurations have close enough latencies, those two setups may actually yield similar real-world performance as well.

Judging by the fact that the KT266 at 1000/100 with PC1600 DDR SDRAM offers noticeably more bandwidth than the MAGiK1 in the same situation we can immediately assume, before looking at the actual tests, that the KT266 does not suffer from the same extreme latency issues as the MAGiK1 with the 100MHz FSB.

As we pointed out the first time we looked at this benchmark, memory write performance is significantly degraded in comparison to read performance because in this case there is actual manipulation of memory (think about it, you can read much faster than you can write). 

The KT266 does fairly well in this benchmark, although the bandwidth it offers isn't nearly as great as the AMD 760 it doesn't suffer nearly as bad as the MAGiK1, which has trouble differentiating itself from the year-old KT133.

Memory Latency - Cachemem

Here's where things get interesting; the KT266 running at 1000/133 with PC2100 DDR SDRAM (synchronous FSB and memory buses, resulting in lowest overall latency) running at a CAS latency of 2, results in a 239 cycle memory latency.  This is still no match for the extremely low latencies the AMD 760 is capable of attaining, possibly because of SuperBypass optimizations in the chipset that help reduce latency, but it is still better than what the MAGiK1 is able to offer.

The different in latency between the KT266 and the MAGiK1 doesn't really take form until you make the switch to the 100MHz FSB, where the MAGiK1 does seem to have some serious issues.

Also notice that the difference in memory latency between the KT266 running at 1000/100 with PC1600 vs. with PC2100 is only 20 cycles.  Combine that with the minimal gains in memory bandwidth noticed earlier and you would be safe to predict that PC2100 won't give you a tangible performance gain over PC1600 when running at 1000/100 (core/FSB frequency) on the KT266.

The fact that the KT133A is still a lower latency solution than the KT266 could hurt the KT266's chances in terms of present day performance since only a handful of applications/benchmarks are memory bandwidth intensive enough to truly stress bandwidth over latency.

Now that we're in the realm of the real world we can see how the KT266 really stands up to the competition.  So far we're expecting it to be slower than the AMD 760 but not by too much, although noticeably faster than the MAGiK1, especially at the 100MHz FSB settings.

For whatever reason, the MAGiK1 seems to perform quite well in the SYSMark suite; the KT266 is not enough to dethrone it as it comes in directly under the AMD 760 in terms of performance.

The standings are what we would expect, pay close attention to the 100MHz FSB performance of the KT266 compared to the MAGiK1.

As a more torturous performance test, we turn to Benchmark Studio Beta 2.0 with a total of 13 stress modules running in the background.  This test can truly bring a system to its knees and it really does help in making the performance differences between the various platforms more pronounced.

Here the KT266 does fairly respectably, although it cannot topple the AMD 760 it does come relatively close although the KT133A was starting to get quite close as well.

The Quake III Arena performance chart is almost identical to the Linpack performance graph we started off the test section with.  The correlation could be the dependency on strong FPU performance and its relationship to memory bandwidth/latency. 

In this case, the KT266 is not enough to live up to the expectations set forth by its older and theoretically less powerful brother, the KT133A.  Again, the AMD 760 continues to dominate.  As we hypothesized at the end of our last Socket-A Chipset Comparison, it is highly unlikely that a 3rd party manufacturer will be able to develop a more efficient DDR SDRAM chipset for the Athlon than the AMD 760 at this point in time.

The standings don't really change as we make the resolution switch to 1024 x 768, although the gap does close down considerably.  This indicates that in the real world, in spite of the inferiority of the KT266 to the AMD 760, you'd only be missing out on a 5% performance advantage under Quake III Arena.  Don't you just love video card limitations?

The performance picture is pretty much the same here, although this time around the MAGiK1 isn't able to creep ahead of the KT266 (1000/133-PC2100).  The more mature KT133A platform is able to offer a 4% advantage over its new sibling, in spite of the KT266's memory bandwidth advantage.  This is the latency vs. bandwidth issue that we were alluding to in the Cachemem performance analysis section.

The real world performance of the AMD 760 in comparison to the KT266 is actually relatively large, on the order of 7%.  UnrealTournament doesn't have a tendency to become video card limited at 1024 x 768 x 32, even with a regular GeForce2 GTS it still scales and varies quite well with different CPUs and platforms.  This allows the performance delta to still remain relatively large between the two chipsets with identical configurations.

The KT133A continues to offer its lower latency as a strong selling point over the newcomer KT266, giving it performance just underneath that of the champion AMD 760.

Expendable continues to reinforce the benchmarks we have seen thus far; the KT266 is what the ALi MAGiK1 should have been, although it is still not able to offer the performance of the AMD 760.

Expendable, being a relatively old game, is not taxing our test bed's GeForce2 Ultra much at all thus allowing the game's performance to be more limited by memory bandwidth, latency and CPU performance.  Since the CPU is held constant, and only the chipset and FSB are adjusted, the performance difference between the KT266 and the AMD 760 at 7% is expected. 

More current titles will show a smaller gap simply because other limitations will come into play, most likely video card related.

Speaking of more current titles, Mercedes-Benz Truck Racing is much more recent than Expendable but still aging in that it isn't the most popular game around and has been out for a decent amount of time.

At 640x480x16, we continue to see what we have been seeing thus far, an 8% performance delta between the AMD 760 and the KT266, with the KT133A falling between the two although its performance is closer to that of the AMD 760. 

Being a much more modern game than Expendable, at 1024x768x32, even our test bed's GeForce2 Ultra is unable to laugh at the game's demands and the AMD760/KT266 performance gap shrinks to under 5%. 

Our final gaming benchmark is Serious Sam, and hopefully after this comparison we will transition to using the final version of the game for all benchmarking purposes but these figures should translate over relatively well into real world performance in the game. 

The KT266 performs relatively well under Serious Sam, approximately 5% away from the AMD 760 and less than 4% away from the KT133A; still far from what we would call impressive.  But if you look at it from the perspective of the motherboard manufacturers, at least it is offering better performance than the MAGiK1, regardless of configuration (100 or 133MHz FSB, PC1600/PC2100 DDR).

The newest game out of the bunch also manages to provide the most limitations for the test bed to play with.  For a theoretical comparison of the chipsets you should always look at the benchmarks with the least amount of limitations (e.g. 640x480x16 benchmarks), but if you're trying to find out how much of a performance difference you'll notice in your day to day usage of a system, often the more limited benchmarks are those that are better suited for this task (e.g. 1024x768x32 numbers).

In this case, the performance difference between the AMD 760 and the KT266 is reduced to just over 3% which isn't bad considering this is VIA's first attempt at a Socket-A DDR chipset and it is already noticeably cheaper than the AMD 760.

Professional OpenGL Performance - SPECviewperf
	Awadvs-04	DRV-07	DX-06	Light-04	MedMCAD-01	ProCDRS-03
ALi MAGiK1 1000/100 (PC100)	69	13	15	4.7	16	10
ALi MAGiK1 1000/100 (PC133)	66	13	15	5.2	16	11
ALi MAGiK1 1000/100 (PC1600)	70	13	16	5.7	16	12
ALi MAGiK1 1000/133 (PC133)	79	16	19	5.9	19	13
ALi MAGiK1 1000/133 (PC2100)	81	18	20	6.7	21	14
AMD 760 1000/100 (PC1600)	79	16	19	6.3	19	14
AMD 760 1000/133 (PC2100)	83	18	21	6.9	21	15
VIA KT133A 1000/100 (PC100)	73	14	17	5.5	17	12
VIA KT133A 1000/100 (PC133)	79	16	19	6.5	19	14
VIA KT133A 1000/133 (PC133)	82	17	21	6.7	20	14
VIA KT266 1000/100 (PC1600)	76	15	18	5.9	18	13
VIA KT266 1000/100 (PC2100)	78	16	19	6.4	19	14
VIA KT266 1000/133 (PC1600)	78	16	18	6.1	19	13
VIA KT266 1000/133 (PC2100)	82	17	20	6.5	20	15

Again we've got a pretty close call here, the KT266 performs ok, although it is unable to topple the AMD 760 and in many cases it is having difficulty distinguishing itself from the KT133A.

Final Words

When the VIA KX133 was launched, with PC133 SDRAM, it was just barely faster than the AMD 750 with PC100 SDRAM. However the improvement in features, availability of the chipset, and the overall cost of the solution were three major driving points behind the AMD 750 to VIA KX133 conversion that occurred in early 2000. After that happened, it would be quite rare that you'd find an AMD 750 motherboard for sale, as almost everything was KX133 based by then.

The AMD 760 is in a bit of a different situation, since the platform does not have many of the performance or feature issues that the AMD 750 had. The only real drawbacks are availability and price, which will unfortunately prevent the AMD 760 from becoming the chipset it could have been. That is a plus for VIA since they clearly have the superior DDR chipset outside of the AMD 760 .

From the perspective of the true enthusiast, the AMD 760 is still and will probably continue to be the highest performing Socket-A chipset platform for a decent amount of time. You shouldn't be too worried about a lack of support for the chipset, especially since the 760MP is just around the corner and will also require much of the same driver support that the current 760 does meaning that there is very little chance of AMD pulling support on the product.

From the motherboard manufacturer's perspective, the AMD 760 is still too expensive of a part to base a large portion of your motherboards on, making the KT266 perfect. We would rather see boards use the KT266 than the ALi MAGiK1, not so much because of the small performance differences that do exist when both are running using PC2100 DDR SDRAM and the 133MHz FSB, but because of what happens when you take the FSB down to 100MHz. The fact of the matter is that the KT266 is quite attractive to motherboard manufacturers, because it is cheaper than the AMD 760, offers performance that is slower than the AMD 760 but not by as great of a degree as the MAGiK1 and it is from a company that has been developing a pretty good track record in recent times.

The KT266 is still a maturing platform, if you plan on becoming an early DDR adopter we'd strongly suggest against going with the KT266 now and sticking with the AMD 760. As you can see, the KT266 is currently not able to really even outperform the KT133A most of the time, but as we mentioned in our Socket-A Chipset Comparison, these chipsets and the motherboards they are used on do mature over time.

We will keep a close eye on the KT266, and revisit it if/when the time comes that we can look at a more mature platform much like what we did in the most recent chipset comparison. Until then, you know what we recommend, and you've got the data to make your own decision so, go for it.

Special thanks to http://www.newegg.com for supplying us with a Retail MSI K7T266 Pro Motherboard