What do you think, is it possible to use dual-processor systems built with AMD CPUs as a basis for advanced home PCs? Can Duron CPUs be used for this purpose? If yes, how efficient they will turn out in configurations like that? Will a dual-processor system prove to be a better choice than a traditional uni-processor one?

We believe, when the ancient dream of AMD to enter the dual-processor market came true this summer, you were also asking yourself all these questions. As you may remember, that time AMD launched its AMD-760MP chipset supporting dual-processor configurations and made its first step towards the hitherto unexplored market of Low-End and mainstream workstations and servers. Those who keep track of the hardware press surely know that Dual-Athlon based systems have proven reliable and stable solutions. That's what gives us the right to treat AMD-760MP as a "Dual success" of AMD.

Most reviews in all sorts of media devoted to the work of dual-processor AMD systems focus on heavy applications, which are typical for servers and workstations. However, we believe that it would be also of great interest to find out how dual-processor mainboards on AMD-760MP feel in home PCs. Indeed, wouldn't it be thriving to see what they are worth in office, graphics and, of course, gaming applications? Perhaps, advanced home users would find a good application for dual-processor configurations as well… Think of the users' passionate response to dual-Celeron systems: what if Duron CPUs worked no worse in pairs?

We're fully aware of the fact that today hardly anyone will buy a dual-processor system on AMD CPUs merely in order to play Quake on it. It would be too expensive, as a mainboard alone costs about $300, then there come the expensive DDR SDRAM modules, a powerful PSU, etc. However, we are pretty sure that as soon as a substantial number of AMD-760MP based mainboards from different manufacturers enter the market, and this way cause a logical price drop, the question of buying a dual-processor system as a home PC will become acute. So, why not to try answering it today?

Closer Look

Till recently the only developer of dual-processor mainboards based on AMD-760MP was TYAN. We guess, there is no use to explain once again why AMD relied on this company at first: its traditionally high-quality products, vast experience in developing elaborate server mainboards and perfect market reputation have definitely been the key factors to determine AMD's partner choice.

The first dual-processor mainboard for AMD CPUs from TYAN became Thunder K7. Following all the rules and traditions of server mainboard development it features an integrated Ultra SCSI controller, a graphics adapter, an Ethernet controller, and is positioned clearly for servers (also rack-mount ones) and perhaps workstations. Some time later TYAN announced a few more new dual-processor boards on AMD-760MP intended for other market niches and featuring a different features sets. In this review we'll highlight one of these mainboards, TYAN Tiger MP.

Tiger MP is the youngest member of TYAN dual-processor mainboards family based on AMD-760MP. It has no integrated stuff: no SCSI, no RAID, no integrated VGA adapter. On the other hand, it makes Tiger MP some $200 cheaper than Thunder K7. That's a kind of mainboard (a cheap solution without any integrated stuff) that is most likely to become a basis for a hardcore home system. In the future, of course.

We'd like to warn you against taking Tiger MP for Thunder K7 with the removed integrated components. That would be incorrect. Tiger MP is an absolutely different mainboard, which was developed anew. Even the layout and location of the main electronic components of these two mainboards have very little in common. These boards are just manufactured by the same company and are built on the same chipset, so this why you might discover some similarities.

Let's cast a keen glance on Tiger MP:

• AMD-762 North Bridge
• AMD-766 South Bridge;
• One AGP 4x slot;
• Four 64/32bit PCI (33MHz) slots;
• Three PCI 32 slots;
• Four 184-pin DIMM slots for Registered PC2100 DDR SDRAM;
• Two ATA/100 channels;
• 4 USB ports;
• Optional RAID-controller;
• Phoenix Plug-n-Play 2MB Flash BIOS;
• ATX form-factor (305mm x 261mm).

Tiger MP doesn't suffer the power supply whims of Thunder K7 (if you remember, the latter has an individual power supply cable for the CPUs). It works well with a regular ATX 300W PSU without any additional voltages.

However, TYAN didn't give up using only Registered memory modules. By the way, it is erroneous to suppose that this requirement is imposed by the chipset: the chipset "doesn't care" at all. It's merely TYAN that has soft spot for Registered DRAM. But there are some rumors in the web saying that systems like that can work normally with non-Registered modules as well, if there are two DIMMs installed, not more. Of course, in this case no one would guarantee high stability and reliability of the system. The specs write it clearly: "Registered DDR only". All the rest is solely on individual enthusiasts' conscience.

The DIMM slots on Thunder K7 were angled, so that the memory modules also stood angled. On Tiger MP the modules stand perpendicular to the mainboard. However, it doesn't prevent the board from being used in rack-mount servers. Many PC companies have already demonstrated their 2U housing solutions based on Tiger MP.

On the board there are two groups of dip-switches, which help to adjust the FSB and memory frequencies: 200MHz, or 266MHz, no "in between" overclocking is possible. In AMD-760MP chipset the engineers implemented synchronous data exchange. It implies that the FSB and memory bus frequencies coincide, i.e. if the CPU works with a 200MHz bus, the memory works at the same 200MHz.

Phoenix BIOS offers a necessary minimum of settings. As you might have already guessed, there are no overclocking facilities, nothing like fine tuning.

Maybe, that was the right thing to do: dual-processor mainboards are not the best place for overclocking experiments. Standard data, time, disks, AGP Aperture Size, input order settings and power saving functions are the only available options.

We can scarcely grumble about the mounting and component layout of Tiger MP. There is a lot of free space on the board and neighboring components don't interfere. The North Bridge is equipped with a pretty huge heatsink, and near the CPU Sockets there are cooler connectors. We have also found some other accessory connectors in other parts of the board.

Test Goals

During this test session we were trying to clear out the following issues:

1. Are AMD dual-processor systems a worthwhile variant for an advanced home computer with its typical tasks?
2. Does it make sense to use a dual-Duron system instead of a dual-Athlon one?
3. Would it be better to buy an uni-processor system based on the fastest CPU instead?

We believe that typical tasks for an advanced home PC are applications for 3D rendering, 2D graphics processing, perhaps CAD, compilers, office packages and, of course, 3D games. Even the most advanced users are sometimes tempted to play some shooter or race game or anything like that, so we assume there is nothing shameful for us to suggest 3D games as one of the tests for the dual-processor Tiger MP. None of them except Quake III have multi-thread optimization, but it doesn't mean that they cannot be run on a dual-processor mainboard. Right on the contrary: we were curious to see whether a second CPU could bring about any performance gain in applications, which are don't know to use its advantages.

The second goal was to find out if Duron CPUs could be used in dual-processor configurations. In fact, there are no technical hindrances configurations like that, because these processors are equipped with all the units necessary for working in a dual mode. The only problem is that the L2 cache of Duron CPUs is two times smaller than that of Athlon processors. And practice shows that L2 cache size is of crucial importance for dual- and multi-processor systems.

Testbed and Methods

When assembling the testbed, we didn't strive for CPUs with the highest clock frequency possible. Naturally, we could easily assemble a system based on two Athlon CPUs working at 1.4GHz. But how would we compare the performances of a dual-Duron and a dual-Athlon system then? There would be no fair competition, since today Duron CPUs haven't gone much further than 1GHz.

Following this logic, we selected CPUs with equal clock frequencies - 900MHz. We carried out four series of tests:

• with one Athlon 900MHz CPU
• with one Duron 900MHz CPU
• with two Athlon 900MHz CPUs
• with two Duron 900MHz CPUs

We ran the same tests with Pentium 4 2GHz + i845 (on a mainboard from Soltek). It was not for nothing that we chose i845 chipset: we believe it will soon become the main figure in mainstream Pentium 4 based systems. As long as we are scrutinizing highs and lows of using AMD based solutions as a home PC, i845 (not i850) is just what we need for the comparison, because i850 is likely to occupy the narrow Hi-End niche of Pentium 4 based solutions.

So, our objective was to obtain clear evidence that would let us assess the performance of uni- and dual-processor systems built on Duron and Athlon CPUs. So that we could see if it were reasonable to pet with duality, or it might be better to decide on a traditional uni-processor configuration.

We used the following software for our tests:

• SiSoft Sandra 2001 (to figure out theoretical performance)
• ZD Winstone 99: Dual Processor Inspection Tests, High-end Business Tests
• WinZiP and WinRAR archivers
• Windows Media Encoder
• 3D Studio MAX, Bryce
• Quake III, 3DMark 2001, Dronez
• Adobe Acrobat

These applications together should cover the whole range of most typical tasks of performance PCs acting as home workstations. For instance, ZD tests include CAD applications, FrontPage, compilers, Photoshop. No recommendations should be given to 3D Studio and Bryce either - these are the most popular applications for 3D rendering, which are intended for users with different skill levels though. Archiving is an everyday matter, while converting documents into the universal PDF format is a common thing for everyone who works with documents. As for games, there is nothing to comment on.

Here is what our testbed looked like:

• Seagate Barracuda III 20GB HDD (7200 RPM, ATA/100);
• 256MB Registered DDR SDRAM;
• AOpen GeForce2 graphics card.

We ran all the tests in Windows 2000 Professional, with DirectX 8, nVidia Detonator XP 21.81 drivers and used NTFS file system.

Performance

Thanks to this synthetic test we can assess theoretical performance of the CPUs and measure how fast the data is transferred between the chipset and the memory. We can see pretty well that a second CPU allows to improve the performance nearly twice. No wonder, as SiSoft Sandra tests the system's ability to cope with a limited number of operations.

Winstone comprises two groups of tests: Dual Processor Inspection and High-End Business Mark. The first group is designed to measure the performance in applications, which are specially optimized for dual-processor systems. These are Microstation MP (a typical CAD application), Adobe PhotoShop and Visual C++. As you have noticed, when an application is "aware" of the second CPU and deliberately works with it, there is no decisive difference between dual-Athlon and dual-Duron systems. Frankly speaking, this phenomenon was a real surprise for us. Who could've expected that Duron CPUs are so smart?

Now let's inquire how these systems feel in applications without multithreading optimization. The results are rather controversial this time: in some applications the performance gain is just negligible, while in the other… uni-processor configurations break ahead of the dual-processor ones! In Microstation SE the lag accounts for good 50%! At first, we simply didn't believe our eyes and ran the tests once again. However, double-checking revealed the same state of things. Later in the article we'll try to explain the reasons for that.

By the by, look at the results of the dual-Duron configuration. Like in case of Dual Processor Inspection tests, it runs shortly behind its Athlon based rival.

According to the developer of Windows Media Encoder, it boasts a code with multithreading optimization. In practice, we observed no difference. The same comment is true for the archiving utilities: WinZIP works equally fast with both systems no matter how many CPUs are used. And WinRAR even seems to prefer uni-processor configurations.

3D Studio MAX traditionally favors AMD processors. Besides, this program is perfectly optimized for dual-processor systems. That's where a CPU tandem should enjoy its greatest success as compared to the uni-processor variant.

Indeed, everything goes right - there is a nearly 100% performance gain. We can witness a dramatic performance gap between dual-Athlon and dual-Duron systems.

In Bryce a second CPU brings about no effect whatsoever. It's clear that this application, being designed generally for a home user, has no optimized code.

We were shocked to see the results in 3D games. There is no logic left here: two CPUs give 100fps, while after one of them is removed the figure climbs up to 120fps… We double-checked the results, tackled drivers and settings, ran the tests in different modes - no way! On the dual-processor AMD-760MP based system games ran slower than on the uni-processor one. A bit later we'll make an attempt to account for this fact.

Subjective Impressions

Well, according to our subjective impressions, all the systems work outrageously fast. Nothing to wonder at again - with 256MB RAM Windows 2K flies like a rocket even if the CPUs are not the fastest. Since we took dual-Athlon 900MHz, there was no reason for a slowdown at all.

Please, mind the fact that dual-processor systems turn out more efficient even when the applications they work with are not optimized for multithread operations. Just try to launch an archiving utility on an uni-processor PC and you'll be unable to launch anything else. If you have a dual-processor configuration, nothing will prevent you from starting archiving and plunging into a MPEG/DVD film at a time. The system will distribute the load equally between the CPUs and you'll feel absolutely comfortable watching the film. This was only one of the many possible examples.

Explaining Weird Results

Finally we've arrived to the most important issue: why in games and some other non-optimized applications dual-processor systems prove amazingly slower than uni-processor ones. Now we'll try our best to explain it, though our assumptions are by no means the final say.

From our point of view, the thing is that all the multi-processor systems, including those built on Intel CPUs, face the problem of cache coherency. To cut a long story short, the trouble is that if some PCI device writes data directly into the memory area, which is currently addressed by the CPU, the processor will have to refresh its cache when writing is complete. Otherwise, the data, which the CPU will read from its own cache, will be outdated. The same thing occurs when a PCI device reads data from the memory and the CPU writes into it. In both cases it's necessary to take care that the cache data is not outdated and to refresh them if needed.

In every modern PC you'll find plenty of devices, which use DMA mode very actively. Among them there are sound cards, graphics adapters, disk drives and so on. If DMA activity is too high, dual-processor systems suffer a considerable downtime, because the cache is to be refreshed too often. In its turn, this leads to a performance drop.

Apparently, the situation is aggravated by an ill-implemented AGP support. As you might've noticed, dual-processor systems fall behind uni-processor ones both at lower and higher resolutions, when the main strain falls upon the graphics card. We can supply no other explanation but bugged AGP and IRQ router drivers. If so, there's some hope that improved drivers will come out soon and will increase the gaming performance.

Conclusion

First of all, we have seen with our own eyes that dual-Athlon systems are stable solutions, which can be applied for a wide range of tasks. This is true for dual-Duron systems, too: as we found out, in most applications there is hardly any difference between dual-Athlon and dual-Duron systems.

To say more, in multithread applications with a specially optimized code dual-processor systems are appreciably faster. Meanwhile, in non-optimized applications dual-processor configurations sometimes turn a lot slower than their uni-processor competitors, no matter whether it's a dual-Duron or dual-Athlon system.

In 3D games the drawbacks of AGP interface come around. Still, we hope that sooner or later these problems will be settled and performance will increase.

Then, using a dual-processor system (based on either Athlon, Duron, or Pentium III CPUs) becomes very much justified if you simultaneously run several tasks, each of them being able to load an uni-processor system completely. We experienced it when watching a DivX film and making some archiving, when instead of gazing at a lazy percentage scale we enjoyed the movie.

The last question is whether it makes sense saving one's finances and assembling a dual-Duron based system. Well, it's up to you to decide, we have no definite answer. On the one hand, the dual-Duron system doesn't fall too greatly behind dual-Athlon one. On the other hand, similarly clocked Duron and Athlon CPUs differ by some $10-20 in price, which is not that much either. So, we guess it would be wiser not to hunt for trifling savings, especially since dual-processor mainboards are "a priori" expensive.