It’s all about choices, something we haven’t had very recently.  When the first 133MHz Pentium IIIs were released, our only viable chipset option was the VIA Apollo Pro 133/133A.  When the Pentium 4 hit the streets, Intel’s 850 chipset was the only available offering.

Luckily today, things have changed considerably.  The table below is a listing of all of the available chipsets for the three major CPU platforms.

Having that many choices isn’t important if the platform is not desirable.  Although the Socket-370 Pentium III/Celeron platform can still service some niche markets (especially those that can’t deal with the incredible heat of today’s higher performance solutions) for the most part, going that route is a dead end.  The Pentium III does not have enough front side bus (FSB) bandwidth or enough scalability to be a good solution going forward. 

The Pentium 4 is becoming increasingly more attractive; with clock speeds rising and price dropping aggressively the processor will eventually find its way into the hands of some enthusiasts.  Unfortunately there isn’t enough platform diversity to drive buyers to the solution just yet.  VIA’s P4X266 chipset was finally released and competing solutions from ALi and SiS will be seen pretty soon as well.  But until the release of Northwood, the Pentium 4 will not be a truly sought-after upgrade path.

This leaves us with the home of the Athlon, Socket-A.  The Socket-A chipset market has become much more than the VIA dominated world it once was.  Competing solutions from ALi and SiS have arrived not to mention that NVIDIA’s first foray into the chipset market will be a Socket-A platform.  With performance high and CPU prices very low, it’s no surprise that more and more users are upgrading to Socket-A systems.  While we have yet to see the BX equivalent in the Socket-A chipset world, a few contenders have come very close.  In this roundup we will help you in your next Socket-A upgrade as well as give you an idea of when you should upgrade.

The Contenders

First things first, there are three notable absences from this roundup: ALi’s MAGiK1, NVIDIA’s nForce and VIA’s new revision of the KT266 chipset. 

The MAGiK1 isn’t present in the roundup simply because by no means is it a viable solution.  As we’ve seen in the past, the solution has some serious performance issues and the cost advantage over competing VIA platforms isn’t great enough to justify even considering it. There have been newer revisions to the chipset which we will take a look at in the future, but for now we'd suggest staying away from it. 

Contrary to popular belief (read: popular rumors), NVIDIA’s nForce isn’t a mess.  NVIDIA has learned a lot in these past few months about the chipset/motherboard markets and their progress has been reflected in the current state of nForce motherboards in Taiwan; they don’t have it perfect yet, but they’re getting there.  The performance of the chipset is still under NDA so we unfortunately cannot bring you benchmarks of it in this roundup but we will be able to soon enough.

As we’ve mentioned countless times before, VIA will be introducing a new revision of the KT266 chipset with a higher performance memory controller.  Some of these memory controller optimizations are already present in the DDR memory controller from the VIA P4X266 chipset.  The chipset will not carry any special name; it will just be a new revision of the KT266.  We should have this in our hands very soon and at that point we will add it to the benchmark mix.

With those exclusions out of the way, let’s have a look at the chipsets we did compare.

AMD’s 760

The world’s first DDR platform is still desired by many.  The AMD 760 chipset was introduced to bring DDR SDRAM to the Athlon processor and it would continue to do so until a third party chipset manufacturer could take over supplying the market with DDR enabled chipsets.  Unfortunately this transition didn’t occur as smoothly as planned, so AMD ended up supplying the 760 chipset for much longer than originally planned. 

We used the ASUS A7M266 for our 760 platform.

SiS’ 735

A highly underestimated entry into the Athlon chipset market, SiS’ 735 chipset has gained a lot of attention recently.  SiS sent out some very high performing reference boards based on the 735 chipset not too long ago, and today we are finally able to find boards available in the retail channels.  What’s surprising is that the performance of the publicly available boards are reminiscent of what SiS claimed with their reference design; the 735 is the real deal.

The 735 chipset has a few characteristics that make it unique.  First of all, it is a single chip solution with the North and South bridges integrated into a single chip.  The two are connected internally by what SiS likes to call their Multi-Threaded I/O Link (MuTIOL).  SiS claims that this link offers up to 1.2GB/s of bandwidth between the North and South bridges; astute readers will realize that there is no need for that much bandwidth between the two bridges but it looks good on paper.  The real benefit of MuTIOL is lower latency operation and superior data arbitration, but for the most part you’re still limited by disk performance when you’re talking about transfers between the North and South bridges. 

Unlike competing manufacturers, SiS has their own fabrication plant where they make their chipsets (for example, VIA chipsets are produced out of TSMC and not directly by VIA).  Theoretically this will keep manufacturing costs low, although it does mean that startup costs are significantly greater than competing firms (constructing a fab plant is not cheap, they can cost close to $2 billion).  A combination of being able to manufacture their own chipsets and a willingness to reduce profit margins in order to gain market share has allowed SiS to sell the 735 chipset at significantly lower levels than competing solutions. 

Click to Enlarge

Although many manufacturers have shown off boards based on the SiS 735, the only manufacturer currently shipping boards is ECS.  We used the ECS K7S5A for our tests.  Because the K7S5A supports both DDR and regular SDRAM, we tested the 735 chipset with both PC2100 DDR and PC133 SDRAM. 

The board itself ran fine; in terms of features, the board is lacking in overclocking features but is priced very attractively at $65 - $75. 

VIA’s KT133A

You should be intimately familiar with VIA’s KT133A chipset.  Only recently have shipments of the KT133A begun to die off, partially because of the extremely attractive price point of DDR SDRAM making DDR solutions the thing to have now. 

We used the ASUS A7V133 as our KT133A test platform.

VIA’s KT266

The KT266 got off to a rocky start but it has finally become a much more mature chipset.  KT266 based motherboards are beginning to replace older KT133A solutions as the boards are finally ready for prime time.

We tested with the MSI K7T266 Pro.

The Test

Windows 2000 Test System
Hardware
CPU(s)	AMD Athlon-C 1.4GHz
Motherboard(s)	ASUS A7M266 (AMD760) ASUS A7V133 (KT133A) ECS K7S5A (SiS735) MSI K7T266 Pro (KT266)
Memory	256MB DDR266 Crucial DDR SDRAM (Micron CAS2) 256MB PC133 Corsair SDRAM (Micron -7E CAS2)
Hard Drive	IBM Deskstar 30GB 75GXP 7200 RPM Ultra ATA/100
CDROM	Phillips 48X
Video Card(s)	NVIDIA GeForce3 64MB DDR
Ethernet	Linksys LNE100TX 100Mbit PCI Ethernet Adapter
Software
Operating System	Windows 2000 Professional Service Pack 2
Video Drivers	NVIDIA Detonator3 v12.41 VIA 4-in-1 4.32V was used for all VIA based boards w/ IDE drivers 580_3012

Note: We tested with NVIDIA's Detonator3 drivers because a majority of this testing was completed before we were given Detonator 4 drivers from NVIDIA.

Memory Bandwidth - Sandra 2001

Since its much better to know why something performs the way it does than just knowing how well it performs, we kick off our benchmarks with some low level memory tests.  These are designed to stress the memory controller and the FSB interface of the platform.  Since the CPU is kept constant, the only variables here are the memory type and the North Bridge itself, making these tests a perfect gauge of what to expect.

The STREAM tests in Sandra 2001 are quite useful in getting an idea of real world peak memory bandwidth figures.  Keeping in mind that very few applications actually demand this much memory bandwidth, let’s look at what the platforms can do.

It is widely known that most implementations of integer code occur in situations where the operating datasets are not very large at all.  This reduces the dependency on high bandwidth memory solutions when dealing with most integer applications.  In the Integer-STREAM test, the SiS 735 takes an early lead with nearly a 10% bandwidth advantage over the AMD 760 chipset.  Keep in mind that both platforms were using PC2100/DDR266 SDRAM.

Take note of the poor bandwidth score from the VIA KT266 platform compared to its older PC133 brother, the KT133A.  We’ll explain that shortly…

Floating Point code on the other hand is generally associated with higher bandwidth datasets, increasing memory bandwidth dependency.  In order to properly simulate this, a more stressful test is used by Sandra 2001 to generate the FP-STREAM scores.  Here the 735’s advantage is cut slightly to 6.5%. 

More noticeable is the KT266 that is able to distance itself from both of the PC133 SDRAM platforms.  In situations where bandwidth is critical (for example, bandwidth intensive floating point applications), the DDR platforms are clearly ahead of their SDR counterparts; on the other hand as we noticed in the Int-STREAM test, in the absence of heavy bandwidth usage DDR SDRAM does not necessarily prevail.

Memory Bandwidth & Latency - Cachemem

The cachemem tests show similar results, although erring much more on the theoretical maximum end of the spectrum rather than somewhat realistic maximums.  Here the difference between the SiS 735 and AMD 760 grows to 14%. 

Writing to main memory takes a lot more work than simply reading from it, hence you get much lower peak bandwidths.  The standings change a bit, with the AMD 760 coming out on top and the SiS 735 equipped with PC133 SDRAM distancing itself from the KT266.  

And now to put sense to it all.  Where memory bandwidth isn’t necessary for high performance, low memory latency is the key to success.  The last time we visited latencies of Socket-A memory controllers we concluded that it would be very tough for VIA to beat AMD in designing such an efficient DDR memory controller.  We cited AMD’s experience with Super Bypass functions and other latency reducing techniques in their chipsets as an advantage that VIA would not have.  When looking only at AMD and VIA, we were right; even shipping KT266 boards are not able to offer lower latency memory accesses than the AMD 760.

However, when you throw the SiS 735 into the fray things change considerably.  The 735 offers 17% lower latency memory accesses than even the AMD 760.  Not to mention the 30% advantage it has over the KT266. 

You’ll also note that the KT266 shows higher latencies than the PC133 equipped KT133A.  This isn’t out of the ordinary as DDR SDRAM does have slightly higher latencies and regular SDRAM.  This is why there is very little performance difference, sometimes in favor of the KT133A, in applications that aren’t memory bandwidth intensive.

Business & Content Creation Performance

As we step into the realm of the real world we realize that higher bandwidth memory solutions are not necessary for simple business application tasks.  The range of scores here covers no more than 2.9 Winstone Marks, or about 5%.  Studies have shown that it’s very difficult to recognize anything less than a 10% difference in performance.  The standings make sense, but they’re not worth discussing since all of the platforms perform just fine.

More memory and a faster hard disk would have more of an impact in business applications than would switching platforms.

ZDM’s definition of Content Creation applications are also not very platform dependent either, with the range here covering almost 4%.  These are simple Photoshop, Dreamweaver and Internet tasks that are once again bottlenecked more by memory size and hard disk performance than anything else.

Office Application & Content Creation Performance

We can always count on SYSMark to make things interesting.

The difference between the slowest platform here and the fastest is just over 10%, meaning that it’s finally noticeable.  In this case, almost all of the DDR platforms perform within a few percentage points of each other while the SDR platforms trail behind.

Again, we have a 10% differential between the fastest platform and the slowest.  This time the 735 takes the advantage.

What all of this adds up to is that the DDR SDRAM platforms, on average, are 6% faster than their regular PC133 SDRAM counterparts in your average office/content creation applications.  Obviously this is just an average, and more bandwidth intensive applications will appreciate DDR SDRAM more while others may appreciate it less.

IT/Enterprise Computing Performance

From fastest to slowest, you’re separated by seven tenths of a second.  When dealing with basic office tasks, you’re going to be hard pressed to find performance differences between these platforms.

Adding some background tasks to the mix helps differentiate a bit, resulting in an 11% performance advantage for the SiS 735 over the KT266 and both SDR platforms.  Only the AMD 760 is able to come within 5%.

Adding even more stressful and bandwidth intensive tasks to the load illustrates the need for more memory bandwidth in truly stressful IT/Enterprise computing environments.  When multiple demanding applications are fighting for a limited set of resources, bandwidth, latency and arbitration are key to high performance; the SiS 735 excels in all of these areas.

3D Rendering & Animation Performance

We have proved countless times before that 3D Studio MAX wasn’t incredibly memory bandwidth intensive.  The range of scores differs only by 1.7%.

The AMD 760 seems to perform very well in all of the SPECviewperf viewsets.  The only plausible explanation being that a more mature AGP controller is working in favor of the 760.  In the future, as more and more geometry data is being sent over the AGP bus to the graphics card mature AGP controllers will become key to 3D performance; especially when it comes to games.

3D Gaming Performance

Most games aren’t incredibly memory bandwidth intensive in that they won’t use gigabytes per second of memory bandwidth, but they are much more demanding than most of the applications we’ve looked at thus far. 

Quake III Arena is very well written and benefits tremendously from efficient memory accesses and low latency operation.  The SiS 735 manages to maintain a 6% advantage here, unfortunately that is only at 640 x 480.  As video card limitations increase (with higher resolutions) that performance advantage will deteriorate. 

In this case, all of the DDR platforms hold approximately a 7 – 10% advantage over the SDR platforms.  DroneZ is a much more up to date engine and is representative of what we can expect from performance in future titles: much more of a dependency on memory bandwidth than what we’re used to currently.

The picture doesn’t change much with AquaMark.

Final Words

This updated roundup was long overdue, but it has actually arrived with great timing.  We are on the verge of seeing NVIDIA’s nForce debut as a Socket-A platform; and courtesy of increased performance pressure by SiS, VIA will be bringing a new revision of their KT266 chipset to market as well.  While the usual recommendation, “wait if you can” holds true here more than ever, there are some out there that are in the middle of upgrading and need an answer now.

With DDR SDRAM so very cheap and the platforms very affordable and mature as well, it only makes sense for you to pursue a DDR solution for your Socket-A system.  The KT133A chipset has served us well, but at this point there’s very little reason to choose it over the competition.  The SiS 735 works fine as a SDR chipset but then again, there are very few reasons to stick with SDR at this point.

For the most part, the three major DDR chipsets perform within a few percentage points of each other.  You can’t really go wrong with any of them.  But obviously that’s not the answer you’re looking for.  There are a few cases, especially in very stressful circumstances, that the SiS 735 truly flexes its muscle.

We are still investigating exactly what SiS did to give the 735 such dominating performance, but it’s definitely very low-level architectural enhancements that make the 735 the solution it is today. 

The real beauty of the 735 isn’t its performance, especially since you’ll find it quite difficult to notice a performance difference (in most cases) between it and the 760/KT266.  No, the real beauty of the 735 is its price.  The ECS board we used in this review retails for less than $80.  We have seen it places for as low as $65 plus shipping.  This makes the board and the platform the perfect companion for the very low cost Duron and Athlon processors.  For less than $200 you can easily upgrade your system to a Duron on a SiS 735 board with DDR SDRAM courtesy of the very aggressive pricing from AMD, SiS and DDR SDRAM manufacturers such as Crucial. 

Even if nForce is about to hit the streets, with a $65 motherboard it’s not too difficult to ditch it after a few months and not feel too guilty.