Introduction

Those of you that are new to the computer hardware scene may not be familiar with the impact Acer Laboratories Inc. (ALi) had on the evolution of AMD processors.  The year was 1997 and AMD, for approximately one month, had the fastest x86 processor on the market.  The K6’s introduction, and also AnandTech’s debut, in April of 1997 marked the first step in AMD’s long road to success.  This was before the Pentium II was introduced, thus giving AMD control of the lime light, and it didn’t take them long to begin to reveal what they were planning for after the K6. 

The successor, codenamed the K6-3D, was to boast improved FPU performance, the weakness of the K6, courtesy of a little known set of instructions that eventually carried the name 3DNow!  However another feature of the, at the time, unreleased K6-3D processor was the fact that it ran on a 100MHz FSB instead of the 66MHz FSB that most Socket-7 CPUs ran at during that period.  Remember that back then, there were no chips with on-die L2 cache, so the FSB frequency not only determined the bandwidth available between the North Bridge and the CPU but also the L2 cache speed.  A 100MHz FSB could bring AMD to within striking distance of the Pentium II, however AMD was in no position to introduce a 100MHz FSB chipset for the Socket-7 platform.  Intel had already abandoned the Socket-7 arena in favor of the new Slot-1 that their flagship Pentium II used.

This is where ALi stepped forward and took the burden of developing a new chipset for a platform Intel deemed dying.  With their Aladdin V chipset, ALi became the first manufacturer to provide a chipset for this new Socket-7 motherboard, which gained the name Super7. 

The Aladdin V beat VIA’s Super7 solution to the streets, and immediately illustrated a 10% performance boost over previous 66MHz FSB chipsets because of the increased bandwidth, even without using AMD’s K6-3D chip. 

Unfortunately, as the K6-3D was eventually released as the K6-2, and as VIA’s solution debuted as well, it became clear that the ALi solution wasn’t the best overall solution for Super7 platforms.  Especially when regarding AGP compatibility with some of the “newer” AGP adapters such as NVIDIA’s TNT2, the Aladdin V chipset had quite a few problems.  As support for the platform died down, more and more manufacturers turned to VIA for their Super7 chipset and ALi slowly faded away from our sights. 

ALi resurfaces at Computex 2000

Half way across the world, in Taipei, Taiwan, ALi returned to the surface after a very long hiatus.  This time around, the treasure ALi brought with them was in the form of the first platform for the AMD Athlon that had official support for Double Data Rate (DDR) SDRAM and the 266MHz FSB (133MHz DDR).

ALi’s presence at Computex 2000 was very important because they were the only company, VIA included, to actually have a DDR Athlon motherboard on the show floor.  While VIA crowded the show with DDR266 banners, ALi actually had a part to show. 

Click to Enlarge

However things weren’t smooth sailing for ALi.  Not only was their part non-functional but out of all of the motherboard manufacturers present at Computex, only a single one indicated that they would be building a board around ALi’s solution.  To top it off, this was one of the smallest motherboard manufacturers in Taiwan, Iwill.  As if things didn’t appear to be gloomy enough, every other manufacturer we talked to indicated that they had no plans to use ALi’s solution nor would they be interested in doing so.  Everyone seemed to be waiting on VIA’s KT266, their DDR offering.

VIA is late to the game: Motherboard manufacturers develop amnesia

Most recently, at Fall Comdex 2000, we saw a completely different picture painted for us by the motherboard manufacturers.  With VIA not planning to debut their DDR solution until next year, and many manufacturers staying away from Intel’s 850 chipset until they have a guaranteed belief that demand will surface, there was a demand for a next-generation chipset to provide customers with what they wanted. 

ALi stepped in and, once again, came to the rescue of the motherboard manufacturers.  While VIA had stated that they would debut a DDR solution for the Pentium III before proceeding to do so for the Athlon, ALi had publicly stated the exact opposite to be the goals of their roadmap.

All of the sudden, the manufacturers that had exclaimed back in June that they wouldn’t come near an ALi solution were featuring their ALi based solutions at Comdex.  Our own Motherboard Editor, Henry Kuo, managed to note a total of 11 motherboards based on ALi’s DDR Athlon solution, including Iwill, who was previously ALi’s lone ranger.

The solution is the ALi MAGiK 1 and it’s the second DDR platform for AMD’s Athlon that we will have taken a look at.  Before proceeding further with this review be sure to read our review of the AMD 760 Chipset which contains quite a bit of information pertinent to the understanding of the benefits of DDR SDRAM and the 266MHz FSB. 

ALi works MAGiK

The MAGiK 1 chipset is split into the M1647, what ALi likes to call the Athlon Super Northbridge and one of ALi’s Southbridge chips, the flagship solution being the M1535D+ Southbridge.

The M1647 Northbridge, for starters, supports both the 200 and 266MHz FSB frequencies for the Athlon’s EV6 bus.  The actual operating frequency of the FSB is 100 or 133MHz, however the bus transfers data twice on every clock yielding the 200/266MHz figures we just stated.

Just as with the AMD 760, the MAGiK 1 chipset requires that the FSB be run synchronously to the memory bus.  Allowing for asynchronous operation with a DDR memory subsystem puts a very heavy burden on the motherboard manufacturers as it isn’t a simple task to accomplish, thus ALi simply limits all memory clocks to be synchronous with the FSB clock.  This means that with a 200MHz (100MHz DDR) FSB, the memory clock must be set to 100MHz as well, and for a 266MHz (133MHz DDR) FSB, the memory clock must run at 133MHz. 

Since the memory bus on the MAGiK 1 is DDR as well, things work out perfectly since the FSB and memory bus operating frequencies end up equaling each other.  This also translates into the exact same amount of bandwidth available to/from the CPU as there is to/from the memory subsystem. 

With two officially supported memory bus frequencies 200/266MHz (100/133MHz DDR) to accompany those supported FSB frequencies, the MAGiK 1 supports the two major types of DDR SDRAM: PC2100 (266MHz) and PC1600 (200MHz).  As the names imply, they offer 2.1GB/s and 1.6GB/s of peak available memory bandwidth respectively. 

As is the case with all other forms of DRAM, there are currently different ratings of these modules available which are differentiated amongst according to their CAS latency.  Currently, PC1600/2100 SDRAM is offered in CAS2 and CAS2.5 varieties, CAS2 being the lower latency and the higher performing solution.  Finding CAS2 PC1600 SDRAM is fairly simple, however finding anything other than CAS2.5 PC2100 SDRAM is currently next to impossible.  The MAGiK 1’s memory controller supports both types of DDR SDRAM, and most motherboards will allow you to control the CAS latency manually which will allow you to, in some cases, to run CAS2.5 DDR SDRAM at CAS2 provided that your memory can handle it. 

Micron PC2100 CAS2.5 DDR SDRAM

Click to Enlarge

Click to Enlarge

While DDR is definitely the name of the game when it comes to high performing Athlon solutions, we have pointed out time and time again that the Athlon platform in general and the Duron in particular isn’t being taken seriously enough in the value market.  This is unfortunate because there is such a large market for AMD here, and it is even more unfortunate because it isn’t directly AMD’s fault that they are in this position, rather it is the lack of any third party chipsets that promote a value platform that is holding AMD back here. 

Although the MAGiK 1 does not feature any sort of integrated video as you would expect to find in a value solution, the M1647’s DRAM controller features SDRAM support in addition to the DDR SDRAM support we just talked about.  This is a very strong feature for system builders and OEMs since they can implement one system design, based around a single motherboard, and simply switch out the type of memory they are planning on using according to the needs of their customers.  Using DDR for the high end and performance market segments, while resorting to conventional single data rate SDRAM for the rest of the market. 

While this won’t thrust ALi’s solution into the sub-$1000 market, it definitely makes it more attractive than AMD’s 760 which only supports DDR SDRAM. 

Obviously with single data rate SDRAM the memory bus can be operated asynchronously from the system bus.  The memory controller integrated into the M1647 Northbridge supports 66/100/133MHz memory bus operation.  The latter two operating frequencies we have had with VIA’s KT133 chipset, however the support for PC66 SDRAM doesn’t make much sense since it would severely cripple the Athlon’s performance.  We find it highly unlikely that a motherboard manufacturer would opt to support PC66 SDRAM on a MAGiK 1 based design.

ALi lists the maximum supported DRAM size as being 3GB on the MAGiK 1, however with most boards shipping with no more than 3 DIMM slots, it is unlikely that motherboard manufacturers will guarantee support for DIMMs in 1GB densities, at least not at first. 

We will also be able to find boards that have both 168-pin SDRAM and 184-pin DDR SDRAM slots on them, allowing for the greatest flexibility however you will not be able to populate both types of slots at once. 

Other than the aforementioned unique characteristics, the MAGiK 1’s M1647 Northbridge offers the same features that we have seen since VIA first introduced the KX133.  The chipset supports AGP 4X, and when using the M1535D+ Southbridge it boasts ATA/100 support as well.  Most, if not all, MAGiK 1 based designs will use this Southbridge.

ALi takes their MAGiK on the road

Along with the MAGiK 1, ALi has released the MobileMAGiK 1 which uses the same M1647 Northbridge as its desktop counterpart but is designed for use with the M1535+ Southbridge.  The M1535+ is simply the mobile counterpart to the M1535D+, their feature sets are identical. 

Obviously this announcement is in anticipation of the mobile Palomino and Morgan cores we talked about in our AMD Roadmap article, however currently ALi is simply demonstrating the chipset used on desktop sized motherboards with mobile components and, of course, desktop CPUs. 

The lone ranger is the first to market

We mentioned at the start of this review that at Computex, only a single motherboard manufacturer was supporting ALi, that being Iwill.  At Computex Iwill already had a board design complete, they were simply waiting on ALi to deliver a stable revision of their chipset.

In spite of the fact that many motherboard manufacturers have jumped on the ALi bandwagon in lieu of the delay of VIA’s Athlon solution, Iwill was still the first to announce support and thus it makes sense that they would be the first to production with a MAGiK1 based board. 

Iwill’s board is the KA266 and should be available shortly. 

Click to Enlarge

Issues

The board we received isn’t ready for public consumption yet with the issues we encountered being due to the relative immaturity of the MAGiK1 chipset.

For starters, there are some compatibility issues with the Radeon and the chipset, however updated AGP drivers should be able to remedy the problems.

The chipset is currently lacking in terms of IDE drivers, however ALi is apparently working on ATA/100 drivers for the Southbridge.  There is no word on an expected release date. 

Iwill supplied us with ATA/66 drivers that improved performance somewhat, however we are still awaiting full ATA/100 support from ALi. 

The KA266 didn’t show any major signs of instability during our testing.  Occasionally the board would fail to warm boot requiring a cold reboot of the test system, and moving from CAS2 down to the CAS2.5 setting in the BIOS would cause instability in 3D games/applications.  We tested with the CAS latency set to 2, however we could not confirm if this was indeed setting the latency to 2. 

The test setup wouldn't complete any runs of SYSMark 2000 which is the reason for the absense of that benchmark, an otherwise normal member to our benchmark suite.

Both of these issues will most likely be solved with the final release of the board, keep in mind that this is strictly an engineering sample that we performed our tests on, so the performance as well as any issues we discussed here can be improved upon between now and the point when the board/chipset is actually in the retail stream. 

The Test

For reference, we have thrown in a Pentium 4 at 1.5GHz while the rest of the benchmarks are run on 1GHz CPUs in order to provide a good clock for clock comparison of the platforms. 

To start off the performance comparison, let’s take a look at Business Winstone 2001.  For those of you that aren’t familiar with it, Business Winstone 2001 is an application based benchmark that measures performance while running a number of complex tasks in business applications such as Microsoft Word and Excel while performing other duties between tasks such as scanning the system for viruses, extracting files from a zip archive, and surfing the net (via locally stored web pages). 

The AMD 760 chipset instantly leaps to the top, unusually enough followed by Intel’s 815 illustrating exactly how mature the 815’s memory controller is, allowing it to perform very well in this test. 

The MAGiK 1, with PC2100 DDR SDRAM comes out around 5% slower than the AMD 760 using the same memory configuration.  Not to mention that the MAGiK 1 with PC1600 DDR SDRAM isn’t performing any better than the KT133 with PC133 SDRAM.

Content Creation Winstone 2001 is much like the Business Winstone benchmark we finished describing, the main difference is that CC Winstone 2001 focuses more on content creation applications instead of business applications.  Examples include Adobe Photoshop and Macromedia Dreamweaver. 

Again, we see that the MAGiK 1 comes out approximately 5% slower than the AMD 760, and when coupled with PC1600 DDR SDRAM it is the worst performer out of the bunch. 

Quake III Arena is an extreme example of the memory/FSB bandwidth dependencies that can be seen in most of today’s first person shooters.  It is because of this quality that the AMD 760 DDR platform did so well in the first place, outperforming the KT133 with PC133 SDRAM by over 10%.

The MAGiK 1 is able to provide a performance boost over the KT133, however it’s not as great as what we saw with AMD’s 760.  In fact, the MAGiK 1 is approximately 5% slower than the AMD 760, both with PC2100 DDR SDRAM.

When you drop the FSB and memory bus frequencies (since they run synchronously) down to 100MHz DDR, resulting in PC1600 DDR SDRAM bandwidth figures, the performance of the MAGiK 1 drops down to below that of the KT133 with regular PC133 SDRAM.

Even with further performance optimizations we don’t expect the MAGiK 1 to be able to pull ahead of the AMD 760, however we can say that there is definitely room for improvement as it should be able to at least perform on-par with the AMD solution.

There isn’t anything seriously wrong with ALi’s AGP drivers as the performance at this memory bandwidth limited setting is similar across the board.  If there were major issues with their AGP drivers we would see some evidence here, and while the slightly below average score may be somewhat of an indication it isn’t enough to worry too much.

MDK2, although played from a third person perspective, shares many of the same characteristics as your everyday first person shooter.  There is a strong focus on textures and not too much of attention is paid to polygon count, making MDK2 a benchmark that will soon be indicative of games of the past, not those of the future where polygon complexity will increase dramatically.

Again, the MAGiK 1 is approximately 5% slower than the AMD 760 when used with PC2100 DDR SDRAM and, once again, when used with PC1600 DDR SDRAM the performance drops to below that of a KT133. 

Once again, MDK2 causes the graphics card, in this case a GeForce2 GTS, to be memory bandwidth limited thus causing the nearly identical performance figures. 

UnrealTournament, another first person shooter, this time using the Direct3D API offered a small performance increase for the AMD 760 over the KT133 when we first looked at the game as a benchmark. 

That small performance increase (around 7%) is trimmed down when the same PC2100 DDR SDRAM is used on the MAGiK 1.  Again, we are seeing the same 5% lower performance on the ALi platform when compared to the AMD 760. 

Using PC1600 DDR SDRAM with the MAGiK 1 chipset doesn’t make much sense as its slower than the KT133.  Although our AMD 760 platform wouldn’t allow for multiplier adjustment, thus preventing us from providing comparable PC1600/200MHz FSB numbers on the AMD 760, we can assume that if the 5% performance differential holds then the AMD 760 with PC1600 DDR SDRAM would be around the performance level of the MAGiK 1 with PC2100 DDR SDRAM, most likely a little slower.

The performance gap closes down as the resolution increases, we still see the same trend however.  The MAGiK 1 is under performing compared to the AMD 760 and when used with PC1600 DDR SDRAM it actually comes out slower than the VIA KT133 using conventional PC133 SDRAM.

Expendable has historically been a great benchmark of memory performance and has always been a good tool for weeding out those chipsets that have sub-par memory performance in comparison to others in their class.

We used it to show that the Apollo Pro 133A had an inferior memory controller than that of the Intel BX and later the i815.  And now the benchmark is illustrating the same thing once again, this time in comparing the MAGiK 1 to the AMD 760.  The performance difference this time is closer to 6%. 

Even in professional OpenGL applications such as those represented by SPECviewperf, the MAGiK 1 continues its 5% performance deficit to the AMD 760.

Using a low level memory benchmark we see that there isn’t a problem with DDR SDRAM providing the bandwidth necessary however there seems to be an issue with the performance of the MAGiK 1’s memory controller, at least in comparison to the AMD 760. 

Even here, the AMD 760 is still 4% faster than the MAGiK 1.

We have excluded the Pentium 4 from this test because it throws the scale completely off making it difficult to see a comparison between any of the other contenders.

The performance situation is much closer here, however equivalent theoretical bandwidth, as we’ve proved throughout this article, doesn’t necessarily translate into equivalent performance. 

Final Words

The AMD 760 now has a companion in ALi's MAGiK 1, unfortunately for ALi, the MAGiK 1 is going to be compared to its more polished peer from now until motherboard manufacturers stop using the 760.

From the perspective of the motherboard manufacturers, the ALi MAGiK 1 is a superior solution to AMD's 760 because of its PC100/PC133 SDRAM support, enabling the construction of custom boards for OEMs and system integrators that desire both SDR and DDR SDRAM support on a board.

However, as our benchmarks have shown, the MAGiK 1 does have a bit of growing to do before it can truly become an alternative to the AMD 760 chipset. With PC2100 DDR SDRAM, the MAGiK 1 falls around 5% short of the AMD 760. While this may not seem like much, once you take into account that the AMD 760, in many cases, is only 10 - 15% faster than the older VIA KT133 with PC133 SDRAM all of the sudden the MAGiK 1 begins to negate the performance benefits of moving to DDR SDRAM.

Using PC1600 DDR SDRAM with the MAGiK 1 is a joke, at least in its current state. The performance level is often times no better than that of a KT133 while costing more.

Luckily there is room for improvement, both with BIOS updates from motherboard manufacturers and with improved driver support from ALi. However by the time the MAGiK 1 is up to speed, VIA's KT266 may be close to hitting the streets if not already making itself a home in what could've been ALi's territory.

At the same time there is also the strong possibility that the MAGiK 1's DRAM controller simply isn't as high performing of a solution as what AMD implemented in their 760. This is a very realistic possibility as the third party manufacturers have never been able to meet the performance levels of Intel/AMD when it came to chipset performance.

If ALi had preceded the AMD 760 launch with the shipment of the MAGiK 1 and had garnered more support months ago, we might have been looking at a much more mature part. Instead we have to play the waiting game and see what the motherboard manufacturers and ALi can do with a bit of "MAGiK."

For more motherboards based on the ALi MAGiK 1 chipset, read our article on ALi's future.