Cyrix has undergone quite a few changes since the release of their first product. Forming from a small number of former Texas Instruments employees in 1988, Cyrix originally did not produce CPUs at all, but rather math coprocessors for 286 and 386 systems. A slow progression into the world of processor soon produced the well received Cyrix 6x86 from older products such as the Cyrix 5x86. The Cyrix 6x86 was a chip that got quite a bit of attention as a result of its good business performance and low price. It seemed that Cyrix was on its way to becoming a major player in the Intel dominated field of CPUs.

This forecast changed, however, with Cyrix's acquisition by National Semiconductor. In 1997 National Semiconductor altered Cyrix's focus, forcing the up and coming chip producer to put their efforts into integrated CPUs. This push brought forth the Cyrix MediaGX chips, chips with integrated sound and video controllers targeted at the extreme budget market.

Next came the highly delayed and very disappointing launch of the Cyrix MII chips. Taking almost a full year to move from a PR-300 to PR-333 speed, it seemed that National Semiconductor was leading Cyrix down a road that would almost certainly result in failure.

Cyrix lay on the brink of destruction in August of 1999 when seemingly out of the blue Taiwanese chip manufacturer VIA swooped in and bought the dying company from National Semiconductor. For the net sum of $167 million, VIA Technologies bought the right to not only the MII processor line but also future Cyrix products. The CPU industry looked gleefully to the future, hoping that VIA could turn the Cyrix line around and make it a competitive product once again.

Cyrix was not the only company to be purchased. Shortly later in September of the same year, VIA Technologies acquired Centaur from IDT. This transaction, costing VIA $51 million, gave them the rights to not only the WinChip processor but also the Centaur x86 microprocessor design. A few of us out there remember the launch of the WinChip processor that promised to bring low cost performance to the budget market. Unfortunately, it seemed that IDT's WinChip processor was destined to fail, as each WinChip processor met criticism from the hardware community and OEMs alike. Once again, it was hoped that VIA Technologies could turn around this seemingly hopeless product and team.

Now holding two of the formerly large budget chip producers, the future of both the Cyrix processor line as well as the WinChip processor line lay in the hands of VIA Technologies.

The first product to arrive from the company that now included both Cyrix as well as IDT was named the Cyrix III. This chip, which arrived in April of 2000, was based off the Cyrix developed Joshua core. The trend of previous Cyrix processors continued, as the Cyrix III with the Joshua core provided respectable application performance at a relatively low cost. Still, enthusiasts and OEMs remained unimpressed. After being abused for a number of years, it was not long before the whole original Cyrix CPU engineering team, leaving the Joshua without much room to grow.

Next came the confusion created by the launch of a new Cyrix III chip, one different from the old. Although the new Cyrix III chip bore the Cyrix name, it was powered by the IDT developed Centaur x86 microprocessor core, code named the Samuel core. Gone were the notorious "PR" ratings and the new Cyrix III featured the same name, a different core, and slower performance than the previous Joshua based Cyrix III. This left many frustrated, as it seemed that Cyrix was certain to go the way of the dodo. People wondered how a company could allow a newer product at higher clock speeds to perform slower than an older one at lower speeds. Cyrix enthusiasts bowed their head in shame: it seemed that the die shrink and decrease in power consumption that came with the Samuel powered Cyrix III were not enough to save the Cyrix name.

But VIA did not give up so easily. Gone are the engineers who developed the original Joshua based Cyrix III, replaced by a hardworking IDT team that would not give up. Their newest creation; another Cyrix III, this time based off a new and improved Samuel core, the Samuel2.

Perhaps, as the old saying goes, third times the charm.

The Samuel2 Core

There is no question that the original Samuel core that powered some Cyrix III chips had some fundamental problems. In order to salvage face, if at all possible, some major changes needed to be implemented in the new Samuel2 core. It had to be cool, have a low power draw, be very cost competitive, and most of all be somewhat powerful.

Many of the aforementioned requirements were met in a manner that one does not typically consider: a fab shrink. Normally this method of reducing cost, decreasing heat, and decreasing price is usually not mentioned since fab shrinks do not come very often. Luckily for VIA, the company was able to start production of the Samuel2 on a 0.15 micron process, making the Samuel2 core the worlds first 0.15 micron x86 CPU.

In terms of manufacturing process, the Samuel2 core is quite revolutionary. Even Intel and AMD's latest offerings are 0.18 micron based, including the newly released Pentium 4. What difference does this smaller manufacturing process make?

By decreasing the size of the transistors in the chip, voltage, current, and thus power were also decreased. By decreasing these elements, both the heat associated with the items as well as the power required to operate the transistors are decreased.

We know heat as an enemy because it decreases the operational range of the transistors inside a processor. By shrinking the process on which the Samuel II is produced, VIA paved the way to higher clock speed Cyrix III processors. It is therefore no surprise that we will see the initial Samuel2 based Cyrix III processors running at 700 MHz and higher. In addition, hitting 1.0 GHz should be no problem with such a small process.

As mentioned before, power consumption is also reduced when manufacturing process is shrunk. This allows the Samuel2 based Cyrix III to further penetrate the ultra portable market where low power consumption is key. The Samuel2 based Cyrix III runs off a low 1.5 volts and requires only 4.5 watts of power for a 600 MHz clock speed and 5.1 watts for the 700 MHz clock speed. This is far lower than Intel solutions. The mobile Celeron 600 MHz (which operates at 1.6 volts), for example, requires 5.8 watts to run. Even a difference of .1 watts is beneficial when it comes to extending battery life.

In addition, the low power consumption of the Samuel2 results in much lower operating temperature. The .15 micron process is to thank for that and does provide the Samuel2 based Cyrix III with an advantage. VIA claims that the chip will perform flawless with only a heatsink. Once again, this allows the Samuel2 core to penetrate ultra portable devices, where space and power requirements do not allow for a cooling fan.

Finally, the price of a chip is reduced, for the most part, when the manufacturing processes is shrunk. Since decreasing the manufacturing process decreases the die size of a chip, less silicon is necessary to produce a chip of an equal transistor count. In the case of the 0.15 micron Samuel2 core the die measures a tiny 52 mm^2, only 69% the size of the original 0.18 micron Samuel and about half the size of Coppermine128 core used in the newest Intel Celeron processors.

More Than Just a Die Shrink

It is true that if the Samuel2 core was simply a die shrink of the Samuel core then chip price could be dramatically reduced, as the chip would have the same complexity on a smaller die. The Samuel2 is more than just a die shrink however, as VIA decided to add an additional 64KB of L2 cache, raising the L2 cache level from 0KB on the original Samuel to 64KB on the Samuel2.

Naturally, the addition of on die L2 cache helps to improve the Samuel2 speed. As we saw with the change of the Celeron core from the original Covington core to the Mendocino and now Coppermine128 cores, the addition of L2 cache can make a big difference. The cache, much like the Duron and Thunderbird's cache, is exclusive, meaning that the L1 cache is not duplicated in the L2 cache. With such a large L1 cache, inclusive cache on the Samuel2 core would essentially render the L2 cache useless as it would constantly be filled with information from the L1.

It is only the addition of the extra 64KB L2 cache and the fab shrink to 0.15 micron that distinguishes the Samuel core from the Samuel2 core. The chip continues to include both MMX and 3Dnow! optimizations found on the original Samuel core and contains the same 12 stage pipeline. The pipeline contains far fewer stages than the Pentium 4's 20 stage pipeline, but does contain the same number of stages as the popular AMD Duron processor.

The bus speed of the Samuel2 remains the same as the bus speed used by the original Samuel core, meaning that the Samuel2 Cyrix III chip will run at both 100 and 133 MHz bus speeds. This is quite a step up from the 66 MHz bus speed on Intel's budget Celeron chip and falls behind the AMD Duron's 100 MHz DDR bus speed (200 MHz effective). This may provide a point of confusion, however, as a 700 MHz Cyrix III running on the 100 MHz bus could actually be slower than the 667 MHz chip that we are looking at today that runs on a 133 MHz bus.

The final slight change that accompanies the new Cyrix III come with the package in which the new Samuel2 based Cyrix III will be delivered. Just like its predecessor, the Samuel2 based Cyrix III will fit in any Socket 370 motherboard and will work in many standard Socket 370 motherboards with a BIOS update. However, this time around the Cyrix III will be available in both a CPGA (ceramic pin grid array) as well as a PPGA (plastic pin grid array) package.

The table below shows a quick summary of the Samuel2 Cyrix III, as well as the original Samuel Cyrix III and competing products from Intel and AMD.

CPU Specification Comparison
	AMD Duron	Intel Celeron					VIA Cyrix III
Core	Spitfire	Mendocino	Coppermine128				Samuel	Samuel2
Clock Speed	600 - 800 MHz	300 - 533 MHz	533 - 800MHz				500 - 667 MHz	667 MHz +
L1 Cache	128KB	32KB					128KB	128KB
L2 Cache	64KB	128KB					0KB	64KB
L2 Cache Speed	core clock	core clock					core clock
L2 Cache bus	64-bit	64-bit	256-bit				N/A	?
System Bus	100 MHz DDR (200 MHz effective) EV6	66 MHz GTL+ / 100 MHz (800 MHz only)					100 - 133 MHz GTL+
Interface	Socket-A	Slot-1 Socket-370	Socket-370				Socket-370
Manufacturing Process	0.18 micron	0.25 micron	0.18 micron				0.18 micron	0.15 micron
Die Size	100mm^2	153mm^2	106mm^2				75mm^2	52mm^2
Transistor Count	25 million	19 million	28 million				11 million	15 million

That pretty much covers the technical side of the Samuel2 incarnation of the Cyrix III. Let's find out if the changes VIA made are enough to raise the Cyrix III from the grave.

The Test

Business Application Performance Explained

It has been a little over a month since we started using Business Winstone 2001 as a standard benchmark in our CPU reviews. Although we gave an overview of our history with the Ziff Davis Media Winstone benchmarks in our Celeron 766 review, we thought that some more explanation of these benchmarks would aid in understanding the scores that are spit out. Therefore, before we get to the actual Business Winstone 2001 numbers, let's see what this benchmark measures and how.

Business Winstone 2001 keeps in track with previous Business Winstone releases as it measures system speed by seeing how fast the system can finish common business tasks. This time around, however, both the applications have changed to reflect current productivity tools and the scoring method has become more accurate.

Long gone in Business Winstone 2001 are outdated applications and utilities. The results of a benchmark performed on an obsolete application provide little or no value for many, as system performance of updated applications are what we are concerned with. With this in mind, it was necessary for Ziff Davis Media to alter the aging Business Winstone 2000 which included older applications such as Microsoft Office 97.

The updated version of Business Winstone, the 2001 version, stresses a system by running the following applications and performing standard intensive tasks on them. The applications included in Business Winstone 2001 are Microsoft Access 2000, Microsoft Excel 2000, Microsoft FrontPage 2000, Microsoft PowerPoint 2000, Microsoft Word 2000, Microsoft Project 98, Lotus Notes R5, NicoMak Winzip, Norton AntiVirus, and Netscape Communicator 4.7. Many of the tasks are performed with other applications open, simulating actual computer use. Some of the routine tasks performed include scanning a temporary directory for viruses, updating database views in Lotus Notes, and compressing various files.

As far as the reporting side goes, we are quite pleased with the way the new Business Winstone 2001 arrives at the final benchmark score. First off, erroneous scores are now eliminated due to the fact that Business Winstone 2001 runs the benchmark five times to produce the overall score. The first run is always thrown out, and the highest score of the four remaining runs is recorded as the system score. Between each run, the system is automatically rebooted and defragmented This method almost always produces scores that are within 3% or less of each other, leaving very little room for variation.

Speaking of scores, how does Business Winstone 2001 figure out what score to give a machine? This is done using a rather simple formula which divides the amount of time required to complete the test by the time required by the "base" machine to complete the test (180,000 milliseconds) and then multiplies that number by 10 (divides the number into ten). Written out, the formula looks like "10 / (total milliseconds from the test PC / total base machine milliseconds)." Therefore, the number given is a representation of how many times faster (or slower) the test system is, times 10. For example, a score of 20 would mean that the system tested performed 2 times faster than the base system.

Now that we understand Business Winstone 2001 a bit more, and know what those numbers it puts out mean, let's see how our current test system based on the Cyrix III with a Samuel II core performed compared to some similar systems.

Business Application Performance - Windows 2000

VIA made it very clear to us that the Cyrix III is targeted at "the needs of a Value PC user" including "productivity application .... and internet application" performance. Since Business Winstone 2001 focuses on such productivity and Internet applications, we were quite disappointed and a bit surprised to see the Samuel2 based Cyrix III perform so low when compared to the competition.

The 667 MHz Samuel2 based Cyrix III, even with its fast 133 MHz bus speed, is no match for even the lowest clocked Coppermine128 based Celeron, the 566 MHz version. The Celeron, which is severely crippled due to its 66 MHz bus speed, easily beats the Samuel2 based Cyrix III, scoring 4.1 points higher or about 20% faster than the Cyrix III.

The slowest Duron available, the 600 MHz model, easily dominates the new Cyrix III as well. This chip, which we have noted many times in the past as one of the best value chips on the market, is able to beat the Samuel2 Cyrix III by 51%.

Content Creation Performance Explained

Another benchmark we recently started using in place of an older version is Ziff Davis Media's Content Creation Winstone 2001. Much like the switch from Business Winstone 2000 to Business Winstone 2001, the new Content Creation Winstone 2001 makes use of updated applications to further stress our test systems.

Rather than use non-intensive applications like Microsoft Word that Business Winstone 2001 uses, Content Creation Winstone 2001 focuses on more demanding applications. Since many demanding applications are involved in making and modifying Internet content.

Content Creation Winstone 2001 uses Adobe Photoshop 5.5, Adobe Premiere 5.1, Macromedia Director 8.0, Macromedia Dreamweaver 3.0, Netscape Navigator 4.73, and Sonic Foundry Sound Forge 4.5 to simulate common tasks involved in making complex Internet content. On the whole, these applications are very intensive, eating up all the power they can get. The benchmark does multitask, switching applications while running others in the background, a fact that requires even more power. The tasks performed by these applications has been modified from the old version as well, simulating actual computer use even better.

Like Business Winstone 2001, Content Creation Winstone 2001 also incorporates the new scoring method, with repeated runs and reboots between each. Also the same is the scoring methodology. Once again the base machine, in this case a Pentium MMX 233 MHz with 64 MB of memory running Windows NT 4.0, scores a 10 in the benchmark. Therefore, the final benchmark score represents how many times faster the tested computer could finish the operations, times 10.

Let's see how our Cyrix III CPU did in this intensive test.

Content Creation Performance - Windows 2000

Once again, we are disappointed with the Samuel2 based Cyrix III's performance. The processor does not seem very capable of handling the intensive Content Creation Winstone 2001 benchmark suite, even though these are the applications likely to be applied in a business situation.

The Cyrix III with the Samuel2 core falls behind the lower clocked and bus speed limited Celeron 566 MHz by 43%. The Duron 600 MHz, which also holds a clock speed disadvantage, manages to dominate the Cyrix III by performing 82% faster.

SYSMark 2000, a benchmark we have been using for quite some time now and runs a variety of applications independently, can give us a view of how a system reacts in the less stressful situation that does not involve multitasking. In the case of the Cyrix III, things do not look good.

The Samuel2 based Cyrix III could not complete the SYSMark 2000 benchmark because the processor would not complete the MetaCreations Bryce 4 test. Bryce 4 is a image creation utility that allows a user to create 3D scenes and landscapes for use in web content and movies. Unfortunately, the Cyrix III consistently hung when the benchmark attempted to use this application.

We can narrow the problem down to the Cyrix III itself, as the same Cyrix III ready motherboard was used to test both the Cyrix III chip as well as the Celeron chip. VIA did note that our sample was "pre-production" but gave no warnings about improper performance. All we can do is hope that this problem is alleviated before the "final" processors hit the shelves or else VIA will have quite a problem on their hands.

Gaming Performance - Quake III Arena - Win98SE

We know that the cache level is fine, as is the clock speed, so what could be holding the new Cyrix III back in 3D gaming? The most likely culprit is one that has been the Achilles heel for the Cyrix and WinChip processors for quite some time: the FPU or floating point unit.

Well, as the Quake III Arena scores show, the FPU of the Samuel2 Cyrix III seems to still be a thorn in the chip's side. Quake III Arena, as well as other 3D intensive applications and games are very FPU intensive. Usually, Quake III Arena can give us a glimpse of a FPU's strength.

In the case of the Samuel2 based Cyrix III, this glimpse is not very promising. Since we know that Quake III Arena is also very memory bandwidth intensive, and we know that the memory bandwidth on the Cyrix III is much larger than that on the Celeron (133 MHz bus versus the Celeron's 66 MHz bus), we can eliminate this as a problem spot. Therefore, the most likely culprit is the Cyrix III's FPU.

Regardless of the problem, the performance of the Cyrix III is sub par. The CPU performs 17% slower than the lower clocked Celeron and 67% slower than the also lower clocked Duron 600 MHz.

Although VIA did suggest that this CPU is not intended for 3D game play, the CPU actually performs closer to the competition in our Quake III test at 640x480x32 then in business centered benchmarks.

At 1024x768x32, the Cyrix III continues to lag behind. Here the CPU performs 15% slower than the Celeron and 49% slower than the Duron. Poor FPU performance or not, the Cyrix III is performing slow considering its 667 MHz clock speed.

Gaming Performance - MDK2 - Win98SE

If it was not clear enough in our Quake III Arena tests, the MDK2 benchmark reinforces the fact that the Cyrix III is not a gamer's friend. At 640x480x32, the Cyrix III 667 MHz performs 29% slower than the Celeron 566 MHz and 75% slower than the Duron 600 MHz.

The results of testing MDK2 at 1024x768x32 remain nearly unchanged from the scores at 640x480x32, suggesting a CPU limitation in all cases but the Duron.

Gaming Performance - UnrealTournament - Win98SE

Unreal Tournament is a very texture intensive game, meaning that the memory bus of a system is placed under considerable stress.

Unfortunately, the Cyrix III's 133 MHz memory bus does not make up for its poor performance. Even when compared to the lower clocked Celeron which is limited to a 66 MHz memory bus, the Celeron still comes out on top. This time the difference is 16%.

The scores remain similar in all cases when going from 640x480x32 to 1024x768x32. Here the Cyrix III is beat by the Celeron by 17%.

Gaming & Professional OpenGL Performance

We typically associate Expendable with memory bandwidth. In fact, it is because Expendable is so memory bandwidth intensive that we include it in our benchmarks. However, the extra memory bandwidth that the Cyrix III has as a result of its 133 MHz FSB when compared to the Celeron does not give it the edge here. We can only attribute this to very poor processor performance that is essentially holding the FSB speeds back.

It is not surprising that the Cyrix III performs the slowest in all the SPEC viewpref tests. As an entry level OpenGL workstation, using the Cyrix III would almost be crazy. Even the Celeron 566 MHz, a chip which we have previously criticized for its poor SPEC viewpref performance, beats the 667 MHz Cyrix III. Professionals dependent on applications such as Pro/Engineer and AutoCAD would only be hurting themselves by choosing the Cyrix III as a workstation CPU.

Anaylizing Performance: Linpack

It seems that on top of the Cyrix III's crippled FPU, first noticed when in Quake III Arena, the Samuel2 based Cyrix III also suffers from very poor cache performance.

As Linpack shows, the access to cache memory is nearly the same speed as the access to the system memory, meaning that the processor's on-die cache is doing almost nothing for performance. For comparison, take a look at the performance of the Duron processor which holds, on paper at least, the identical cache layout and design. The Duron performs much faster when accessing memory from the cache, while the Cyrix III is left in the dust.

We can not determine what this is due to simply because we do not have enough information on the processor's architecture. What we do know, however, is that it is this poor cache performance that is most likely limiting the speed of the Samuel2 based Cyrix III in every situation.

Conclusion

From the looks of our Samuel2 based Cyrix III processor, it may be a while before the Cyrix line can achieve success again. There is almost no question that as a desktop solution the Samuel2 based Cyrix III is destined to fail. This is not only due to its poor performance but also its cost. Our sources say that the Samuel2 based Cyrix III is expected to cost $20 less than a Celeron of equivalent speed but can not be confirmed.

If proved true, this means that at the estimated release speed of 700 MHz, the Cyrix III will cost about $56 (based on today's prices). For $6 more, one can get a Celeron 566 MHz, like the one tested in this review, that outperformed the Cyrix III in every test. Since the same motherboard and system components can be used, no extra cost will go into building this Celeron system. In fact, it would be cheaper to build the Celeron system simply due to the fact that the memory does not need to operate at 133 MHz (like it does in the Cyrix III) but only needs to be rated to 66 MHz.

To put the final nail in the coffin, all Duron processors at and below the speed of 700 MHz can currently be found for less than the $56 price that we estimated for the Cyrix III. The Duron 600 MHz that grossly outperformed the Cyrix III in our benchmarks can actually be found for $7 less than the price we estimate the Cyrix III at. Although the current budget solutions for the socket-A Duron are not as powerful as the budget solutions that the Celeron has for it, there is really no question that even these low performing budget solutions will easily outperform the Cyrix III.

What part of the market does this leave the Cyrix III? Well, with the budget market essentially eliminated all the Cyrix III has left to cling to is the ultra portable market, where the low power consumption of the chip may prove to be attractive. It is unclear how the Cyrix III will perform when compared to other "ultra portable" CPUs, but it is necessary to keep in mind that VIA is pushing the Cyrix III into the value PC market, not the Internet appliance or handheld market.

With this in mind, we can only say that the third incarnation of the Cyrix III, the one based on the Samuel2 core, is destined to go the way of recent Cyrix products. In order for VIA to reach their goal of 10% market saturation, they will either have to drastically change the Cyrix III as it currently is or come out with a new product that is priced lower and performs better. The additional 64KB of L2 cache and the shrink to the 0.15 micron process do not seem to be not enough to save the Cyrix III from going the way of the dodo.