Usually, when we're bored, we find something to occupy our time.  Whether this is a method of fooling ourselves into thinking we're being productive (when all we're actually accomplishing is watching our favorite TV show or schooling a few boys in the ways of Quake), or a way to occupy our time and make sure we don't turn our attention to anything destructive (that telephone is looking mighty overclockable right about now...), it does happen.  Assuming that the latter is true, the best case in point would have to be AMD with the release of their K6-2 350, a product intended to occupy a portion of time, living off of the short attention span of all tweakers when it comes to high performing hardware.

AMD last left off with the disappointing introduction of the K6-2 333, running on a 95MHz FSB and only performing around 1% faster than a K6-2 300 in most applications due to the decrease in FSB (100 > 95MHz) frequency.  With both chips equally overclockable to 350MHz (for the most part), the 333 was a pure bomb among those who were a little more knowledgeable about the value of a 300MHz part.  At the same time, Intel was reaping in the benefits of a low cost, high performance processor solution for the masses, the Celeron A.  The need for a new chip from AMD was obvious, and with the release of the next generation K6-2 processors in late November, there was no time to wait around.  So in essence, to fill the gap between the yet to be released K6-2 400, and the average performing K6-2 333 which was being overshadowed by Intel;s Celeron A, AMD needed something to do to keep them occupied...their task?  Constructing a 350MHz K6-2.

The K6-2 & 3DNow!

You have probably heard many "techno-geeks" complain about how we don't really need all of this power in our computers today, ever wonder why that topic comes up so frequently?   Consider this, your CPU has the power to process millions of commands in a single second, but how much of your CPU's full potential do you imagine is being completely utilized by simply clicking on the Start Menu?  Or in a more related sense, how much of your CPU's full potential is being used every time you run Quake 2?  The answer is obviously a very limited portion, while there are many interactions taking place between your CPU and on-board cache subsystems, system memory, your video bus, and other such things, your CPU itself, more specifically the FPU is only performing a limited amount of tasks in a highly repetitive manner. 

On the hardware side of things, the FPU of a processor is designed to basically handle anything that comes its way, in contrast, on the software side, a 3D game is designed to take all it needs from your system to run and leave it at that.  Whether or not your computer has a FPU capable of calculating a trip to Mars does not matter to a game that simply requires the execution of a few relatively simple FPU instructions.  This is where AMD saw a chance to make an impression that would definitely last, and more importantly a chance to make an impression that would open a window of opportunity for future generations of AMD processors to come.   Instead of concentrating on re-designing the FPU of the K6, why not simply offset the weakness by using a separate set of hard-coded instructions (instructions built into the CPU itself vs software based instructions) to perform common tasks used in 3D games.   This was the hand AMD brought to the table with the K6-2, a winning hand, that is if AMD could pull it off.

The AMD 3DNow! Instructions often fall into the MMX ditch as far as comparisons go, and it should be mentioned here that unlike Intel's 57 MMX Instructions, the 3DNow! instructions actually serve a real world, noticeable, gaming purpose.  It is true that MMX instructions do help in some cases, however the hype surrounding the integration of MMX instructions into the Pentium line of processors blew the realistic impact of the instructions completely out of proportion.   3DNow! exploits a technology known as Single Instruction Multiple Data execution, or SIMD for short.  What this technology does is, as the name implies, it applies a single command (or instruction) to multiple sets of data simultaneously.  A real life example of something like SIMD execution would be much like picking berries.  Instead of picking a berry off of a bush, washing it, then returning to the bush to collect more you could simply collect a handful of berries at once and retrieve them simultaneously.   The key to understanding SIMD is the simultaneous execution of the instruction, therefore saving time.  SIMD is not the only benefit the K6-2 receives from AMD's 3DNow! instructions, however it is definitely a plus.  If you have ever written any code that requires heavy floating point calculations you will know that floating point division is an extremely slow process from the perspective of the CPU.  AMD saw this as a major problem and integrated a reciprocal multiplication function into their 3DNow! instruction set.  For example, instead of dividing a number by 4, using reciprocal multiplication you can achieve the same result by simply multiplying by the fraction 1/4.   While this may not seem like much, you must understand that a processor can multiply faster than it can divide, in this case, instead of dividing by 4 you are multiplying by 1/4, which can make a world of difference in terms of performance on a more complex level.  Since the reciprocal multiplication function is hard-coded on the K6-2 it actually provides a healthy performance increase in many situations.  It is instructions like these that make the 3DNow! instruction set different from the old MMX instructions that drove the market wild what seems like so long ago. 

Parting the Waters

With that brief overview of the advantages a K6-2 offers over its older brother, the K6, let's take a look at the requirements of the K6-2 350 as well as the differences between it and its predecessors:

AMD K6-2 System Requirements
•	Super7 motherboard with official support for the 100MHz FSB
•	Motherboard support for a 2.2v core voltage and 3.3v I/O voltage
•	BIOS support for the K6-2 350
•	Timing patch for Windows 95 users (available at www.amd.com - not for Win98 users)
•	x86 operating system environment

Aside from the increased clock frequency, the K6-2 350 offers very little over the 300, or 333.  In terms of performance, the 350 is around 4% faster than the K6-2 333 in business applications, and approximately 2% faster than the 333 in gaming situations. 

While the figures seem relatively low, this is a normally clocked comparison, once you migrate to the field of overclocking the picture changes by a more noticeable degree.  While the 300 and 333MHz chips normally topped out at 350MHz in terms of a reliable overclock, the K6-2 350 manages to push the limits even further with success stories popping up in the overclocking community of 350MHz chips running at up to 450MHz with minor cooling enhancements, such as larger heatsink/fan combos, or even peltier coolers. 

The 350MHz chip AnandTech received for evaluation unfortunately only made it up to 400MHz using normal cooling techniques on the FIC PA-2013 MVP3 test bed.  In order to achieve reliable performance at 400MHz, the K6-2's core voltage had to be increased to 2.4v from the standard 2.2v setting.   According to AMD, the K6-2 can handle core voltages up to 2.5v without truly harming the chip, however that does not take into account any type of harm you are causing to the CPU by overclocking it.  Clocking the chip at 450MHz resulted in a frozen system at the power on self-test (POST) and failed to boot successfully.  Using an experimental cooling technology similar to those used by Kryotech to thermally accelerate the K6-2, AnandTech managed to get the processor to boot at 450MHz and run through a set of gaming performance tests.  The Winstone 98 (business applications) tests took too long to complete and would not run through reliably enough to obtain a score.  Overall, you can expect a fairly decent percentage of 350's to be able to make it up to 400MHz, with the cream of the crop being able to hit the 450MHz marker.  Is 500MHz a possibility?  It doesn't seem so.  Even after cooling the processor down to well below freezing, the system wouldn't even boot past 450MHz.

The Test

The Socket-7/Super7 Test System Configuration was as follows:

• AMD K6 233, AMD K6-2 333, AMD K6-2 350, Cyrix 6x86MX PR200+, Cyrix M-II 300, Intel Pentium MMX 233

• FIC VA-503+ Motherboard

• 64MB PC100 SDRAM

• Western Digital Caviar AC35100 - UltraATA

• Matrox Millennium II AGP Video Card (4MB)

• Diamond Monster 3D-2 Voodoo2 Graphics Accelerator (12MB)

The Pentium II comparison system differed only in terms of the processor and motherboard in which case the following components were used:

• Intel Celeron 266/300, Intel Pentium II 400

• ABIT BX6 Pentium II BX Motherboard

The following drivers were common to both test systems:

• MGA Millennium II Drivers v4.07.00.700

• DirectX 6 SDK

• Quake 2 v3.17 w/ 3DNow! patch

All tests were run at 800 x 600 x 16-bit color

On the business end of things the above graph shows the obvious trend towards generally more efficient integer calculations and cache utilization in today's processors.  The K6-2 falls a tad short from reaching the performance point set by the 400MHz Pentium II, however it offers a determinable performance improvement over older Pentium II's.  What you must keep in mind is that the K6-2's L2 cache will always remain at or around 100MHz, while the L2 cache of the Pentium II is constantly increasing with clock speed.  Take the Pentium II 300 for example, it's L2 cache runs at 150MHz (50% of its clock speed), while the K6-2 300's external L2 cache runs at 100MHz (100% FSB speed).  If we compare a Pentium II 300 to a K6-2 300, the K6-2 300 comes out on top due to the advancements it offers over the Pentium II.  However, by bumping the clock speed of the Pentium II up to 400MHz, therefore increasing the L2 cache speed to an incredible 200MHz (50% clock), Intel's flagship begins to outpace the K6-2 whose L2 cache remains at 100MHz.  It is for this reason that the life of the K6-2 is being limited, and it is for this reason that AMD has planned much more than a clock speed increase for their 400MHz K6-2.  While the release information is shady at best, there will be an increased number of core optimizations for the K6-2 once it hits the 400MHz barrier.  We can expect the performance of the K6-2 400, in business applications at least, to out pace that of the Pentium II 400 by a determinable, albeit unnoticeable in real world situations, amount.

The 3DNow! instructions are clearly taking their toll out on the frames of Quake 2 as the 3DNow! enabled K6-2 offers Pentium II-class levels of performance at a considerably lower cost.  Unfortunately, if you take that 3DNow! optimization away from the K6-2, you get performance no better than that of an original K6 running at a higher clock speed.  This is the case in quite a few games which do not make adequate (if any) use of AMD's highly publicized 3DNow! instructions.    

The Future 

If you've already set your mind on a Super7 Motherboard + K6-2 combo, the K6-2 350's $30 price increase over the 333 is well justified if you factor in the possibility of a successful overclock to 450MHz.  Don't get your hopes up though, the limited publication of astounding success in this area should keep your minds a bit more than skeptical.

As far as being an upgrade from scratch, the recommendation here clearly goes to the ever-so-popular, wait another month and see.   AMD is planning something with their K6-2 400 which should, both, drive prices even lower for the normal K6-2 processors, and give you a more enticing offer for a Pentium II equivalent at a lower cost. 

Die hard gamers will want to stay away from the Super7 platform as AMD has yet to summon the support they promised for their 3DNow! instruction set, what good is 60 fps in Quake if 30 fps in Half-Life is all you can expect?  The Celeron A offers an equal price, for the standard in gaming performance when it comes to Pentium II-class systems, and for the true spender, the Pentium II is still at the top of the money tree. 

The K6-2 350 ends up being a wonderful processor for a high-end, low-cost, system, especially with the price of decent Super7 motherboards ranking in around the $90 level.  As the heart for a < $1500 system, the K6-2 350 does everything you could possibly ask for, and more, however as the heart for a $2500+ system...I'm sorry to say, AMD had better crank that K6-3 out soon, as Intel comes out on top in that market.