The AMD K6-2 is a very good overclocker. Probably only the Intel Celeron is a better overclocker. This makes the chip even more interesting. Already in non-overclocked state it delivers Pentium II level performance for much lower prices. The K6-2 has unlocked multipliers so is not like Intel Slot 1 CPU's bound to one multiplier. This gives you much more freedom, safety since you don't need to adjust the FSB, and bigger chances at success than with most Pentium II systems. AMD didn't have all the problems with relabeling CPU's as Intel did and was also aiming at upgraders who were still using the 66MHz FSB. However to release the real potential of this chip it requires much more than Intel Slot 1 CPU's a good and stable motherboard.

Some common questions

Is overclocking dangerous?

More and more people overclock their PC. Many of them don't seem to realise that overclocking always has a certain risk. Overclocking is nothing more than pushing your hardware over the limits of its design. With other words it's going outside the specs of hardware. Officially your vendor warranty becomes void when overclocking. So when you overclock your PC, proper functioning, or even functioning at all, is not guaranteed. Overclocking can damage your PC hardware. If not in short term it's very likely to decrease the lifetime of your hardware in the long run. More important even is that overclocking can result in unstable systems, which could result in crashes, data corruption and data losses.

By looking around on the web you might get the impression that these risks are very small. But then ask yourself for instance the question if a 300 MHz CPU runs always so stable at 333 MHz, then why isn't it sold as a 333 MHz CPU? So overclocking is always a risk and never a guarantee. This is no matter what kind of hardware you have. So when overclocking always keep this in mind.

When is overclocking a success?

This is the most important question. May people don't realise that successful overclocking is more then set you CPU to higher speed and see if it boots in Windows. Overclocking is successful when the result is stable system. But when is system stable? I once heard someone tell on usenet that he overclocked it's Celeron 300A to 450Mhz. He claimed that it hadn't crashed for five hours since he overclocked so his conclusion was that is was stable. Well I wouldn't call that stable yet. Stability is something you find out after many hours trying. Another example is yet another Celeron 300A user how claimed it's CPU was stable at 450 too. It "was perfectly stable and only crashes after 30 minutes when playing UnReal." Well, that's not stable.

Let me illustrate this by my own system. I owned an AMD K6-2 300 MHz with an ASUS P5A mainboard. Overclocking was quite easy and without any special adjustments I got it at 350 MHz. I could run Windows 98 without problems, run Final Reality or play Quake II for hours. I could have concluded that the system was stable. But as Windows NT users know NT is pickier on the CPU than Windows 98. Although my system booted Windows NT without problems and I could run all my applications (mainly Internet stuff, 3D tools and Visual C++) without problems. However I got about once or twice a week a Blue Screen of Death without any visible cause. Clocking my CPU back at 300 made the crashes disappear and I could run NT for weeks without crashes. Even more picky about the CPU then Windows NT is UnReal. Although I could play Quake II for hours UnReal crashed without any notice within a few minutes. I now have a K6-2 350 @ 380 and the same situation applies.

So this illustrates my two points: stability is only proven after days of operation and stability can differ from application to application. To make thing even more complicated one must realise that some non-overclocked systems also crash once in a while. So stable is something that depends on the applications and can only be determents after a few days a heavy use.

Requirements

Cooling

The K6-2 CPU is a very cool CPU thanks to the .25 micron-technology. No special peltier-coolers are needed. Just get a decent fan with a big heatsink. A little bit of thermal paste can increase the effects of your heatsink very much. The thermal past makes heat transfer between the CPU and heatsink better by improving contact. However remember that the heatsink itself transfers heat much better than any past, so use as little as possible; just enough to get a smooth contact.

There are special AMD coolers available with special contact areas of a different material then the rest of the heatsink. They are not necessary - I own one myself - since they don't cool better (or worse).

Another interesting point is software cooling. It cools the CPU down by halting some parts of the CPU at moments that they are not used. Operating systems like Linux and NT have native support for these so called HLT instructions. For Windows 9x extra software, like Waterfall and CPUidle, is required. On one hand I'm very enthusiastic about these software since they really cool your CPU a few degrees. On the other hand cooling only occurs when parts of the CPU are unused. So, for instance, when you are playing games software cooling can NOT cool your CPU. So when overclocking software cooling does NOT increase the chance at success and it will not prevent your system from overheating, so it's no solution for overclocking problems.

Motherboard

For overclocking the following things are important.

• Core-Voltage: choose one supporting 2.1V, 2.3V, 2.4V CPU core voltages. The ASUS P5A is for instance an excellent choice since it lets you pick any voltage between 2.0 and 3.5 for the core.
• Front Side Bus (FSB): choose one that has besides the 95 and 100 MHz options also settings like 105, 110, 112, 115, etc. The more the better because the more options you have.
• Chipset: at this moment there are a few possible choices. I would not recommend an Intel TX since it doesn't support the 100 MHz FSB (officially 66 MHz - unofficially 83 MHz). Same story for the VIA VP3 (officially up to 66 MHz - unofficially 83 MHz) and SiS 5591 (officially 99 MHz - unofficially 100 MHz). So basically you should choose between the VIA MVP3 and the ALi Aladdin V.
  The MVP3 has the benefit that it often let's you run your memory at AGP speed (normally 66 MHz) instead of bus-speed (the so called a-synchronic memory setting). So for those who upgrade and don't want to buy new PC100 memory I could recommend a board with the MVP3 chipset.
  The Aladdin V chipset, like used in my personal favourite the ASUS P5A, has one important bad item: it can only cache up to 128 MB of memory when equipped with 512 kB of cache. For most users however 128MB is more than enough. Boards with the Aladdin V chipset are most of the time a bit faster then equivalent MVP3 boards.
• Cache: the more the better. Although ALi Aladdin V boards with 512kB often outperforms a MVP3 board with 512kB cache, there are very few Aladdin V boards with more than 512kB cache.

Memory

Get PC100 memory: it costs hardly more then normal PC66 SDRAM, and besides the 3% performance increase compared to PC66 memory, you can run at 100MHz bus and higher. One important remark must be made. There are different types of PC66 and PC100 memory. PC100 memory only means that it's suited for the 100 MHz bus. That means it is 8ns or faster. For PC66 counts that it is in the range of 10ns or slower. Note that a 133 MHz bus can NEVER be achieved with 8ns memory, which most PC100 memory is, since 1/8ns = 125 MHz. This also explains why some PC66 memory modules can be used with a 100 MHz FSB; 1/10ns = 100MHz.

Tips, Tricks and important Facts

FSB bus

Increasing the FSB improves performance more than just overclocking your PC but brings also greater risks. The FSB bus has effect at 5 points:

• L2 cache: runs at FSB on Super Socket 7 boards. So increasing your FSB will also increase performance of the L2 cache. On the other hand L2 cache can be the reason for instability when overclocking. When your system is instable after overclocking disable the L2 cache in the BIOS. If your systems is stable (and slow) at that point the L2 cache cannot handle the higher bus-speed.
• Memory: runs at FSB too. A higher FSB will increase performance quite a bit but can result in memory crashes. If your system is not stable after overclocking you could set in the BIOS the following settings:
• SDRAM CAS Latency: from 2 to 3
• SDRAM RAS to CAS delay: from 2 to 3
• SDRAM CAS Precharge time: from 2 to 3
This decreases memory-performance a bit but could make your system stable again.
If your board has the asynchronic memory setting (memory at AGP speed) you could also try that, but often, even if your system is stable, it will then be slower than before overclocking.
• HDD Performance: is directly related to FSB. Many people don't know that your IDE controller will run at the same speed as your PCI bus. If you experience instability after overclocking you could try to disable UDMA in the BIOS. This often decreases the effects of electromagnetic noise. When you're only having one device on a controller it won't even impact performance since UDMA is not faster but only has a higher throughput. You could also try to lower the PIO mode of your HDD. However this significantly slows your HDD down undoing all the effects of overclocking.
• PCI bus: is related to the FSB. Many cards, like modems and soundcards will not benefit from a higher PCI bus. Other cards like Voodoo cards or SCSI cards can get a big performance boost. With SCSI cards you should be very careful since instability there results directly in data losses etc. Videocards nowadays can handle a bus up to 37,5 MHz so no problems should arise there.
• AGP bus: is also related to the FSB. However AGP cards don't benefit much from a higher bus speed. Overclocking your video memory with programs like Matrox Overclock will let you gain more. Note that most AGP cards will not function correct at AGP bus speeds higher than 75 MHz and can even be seriously damaged!

The following table will show the relation between FSB and PCI/AGP bus speed.

Note that these values are not absolute. They can also differ from mainboard to mainboard, so always check your manual. The green values are the standard values. The red value's are the values that are the most dangerous; more likely to be unsuccessful and also more likely to damage your hardware. The values in blue are one of the reasons why I like the ASUS P5A so much. The 83 bus is not dangerous at all, since it doesn't overclock the PCI and AGP bus. This can be very helpful when overclocking the K6-2 266 MHz and 366 MHz or other CPU's like Intel Pentium, IBM/Cyrix 6x86MX, AMD K6, etc.

Core-voltage

When overclocking little leaks, noise, etc. increase and can result in instability. Increasing the core-voltage sometimes compensates for this. The K6-2 is guaranteed to work between 2.1 and 2.3V and the limits are 2.0 and 2.4V. Many K6-2 overclockers have reported that setting the voltage to 2.3V often helps. But always try at first at plain 2.2V. I have my 350 running at 380 and had my 300 running at 350 without the need to increase voltage.

Software

• AMD K6-2 Patch: There's a bug in Windows 95 which makes Windows 95 crash sometimes during boot in combination with K6-2 350 MHz CPU's. To correct this problem download the patch for OSR2 and higher to fix this. I don't know if this happens with overclocked systems too, but you could install the patch anyway.
• 3DNow optimisations: Well this has actually nothing to do with overclocking but can give your system even bigger performance boosts. Besides installing DirectX 6 which has 3DNow support you can also find all kinds of patches at www.3dnow.net and of course AMD's own list.

Success-rates and Performance gain

Success?

What are the chances at success? Well this depends on the type of mainboard, the extra PCI cards you have, your HDD and of course the CPU itself. No two CPU's are completely identical so success can differ from individual CPU to CPU. However the two "turbo-settings" 103 MHz and 105 MHz FSB, which are only a little bit above official settings, are very likely to work when using quality hardware. The higher the FSB, the smaller the chance at success and the bigger the chances at instability.

The following table (based at sources from the Internet and personal experience) gives you an impression of what is possible.

	Good chances	If you’re Lucky	Record
266	300 (3x100)	315 (3x105)	350 (3,5x100)
300	333 (3,5x95)	350 (3,5x100)	360 (4x115)
333	350 (3,5x100)	380 (4,5x95)	380 (4,5x95)
350	380 (4,5x95)	400 (4x100)	450 (4,5x100)

Table 2, expectations when overclocking

As you can see one setting higher is often possible. Note the blue-item. From my experience getting a 300 to run at 350 at a P5A isn't that hard. I have tried several systems and without exception I got em all at 350 on an ASUS P5A without even adjusting any voltage setting. I'm not claiming that every K6-2 300 will run at 350 on this mainboard, but chances are pretty big.

The ideal setting

I would personally recommend not going too far from the standard FSB settings since the chances of success are the best there. Therefore I would for instance rather run a 300 MHz CPU at 333 (3,5x95) than 315 (3x105). This although benchmarks often show about the same scores for both settings. In short one could say that for business applications you could best get a highest possible FSB. With business applications you're CPU is often not the limiting factor but memory, cache, SCSI cards etc. are more important. They benefit very much from a higher FSB. When playing games the pure CPU/FPU power is important. That means getting the CPU at the highest possible clock. Memory timing and cache are less important.

An exception is when using PCI video-devices like PCI SVGA cards or a Voodoo card. Since the PCI bus scales with the FSB you could best choose a high FSB. Also the real potential of the 3DNow! instructions can only be fully used when the 3DNow registers keep filled. This requires a fast L2 cache. For instance Quake II will benefit very much from a fast L2 cache. So when playing Quake go for the high FSB, especially when you own a Voodoo card.

The Gain!

That of course is what it's all about. The following benchmarks give you an impression of what I got when using several CPU's. (ASUS P5A, AMD K6-2 300, 128MB SDRAM PC100, Matrox Millennium G200 8MB AGP, Diamond Monster II 12MB, DirectX 6, 3DNow! patch.)

Figure 1, business performance with the AMD K6-2


Figure 2, gaming (Quake II) performance with the AMD K6-2

With Quake the benefits of the L2 cache come forward very much. The 315 setting wins it from the 333 setting. But in the business test the 315 also comes close to the 333 setting. In both situations however the 350 setting wins the grand price. For the 300 I would always recommend this setting since it's a relative safe setting and, besides when using real high FSB combinations, also the fastest.

For the 333 MHz and 350 MHz similar settings exist. For the 333 the 350 and 380 setting and for the 350 the 380 and 400 setting. Those settings are relatively safe and many users have reported getting their system stable at this speed.

Future

AMD has just released three new K6-2 CPU's; the 366, 380 and 400. The new released 366 and 380 K6-2 CPU's are not more than higher-clocked of the current models. The 366 has just like the 266 a 66 MHz FSB. This will make the CPU mostly attractive to laptop-users and upgraders from older mainboards. The AMD K6-2 400 however has a new core. This core, which also has an improved FPU, will need a new BIOS upgrade to fully exploit it's potential or even function at all. The first tests show that this new core indeed increases performance even more. As soon as I have some test-results I'll let you all know.

The K6-2 400 core actually is the same as the core of the announced K6-3. The K6-3 will have it's own L2 cache running at full clockspeed. The L2 cache on the mainboard will then become L3 cache. This brings also new overclock-possibilities. You can now disable the L2 cache of the mainboard in case of instability without losing much performance; about 4% according to careful estimations based on Alpha systems where L3 cache is not uncommon. First test also show that the K6-2 400 is very good overclockable so the future looks bright. We might get even more value for the buck.

Written by A.A.Gerritsen
for the CPU Site,
December '98

All Names and Trademarks on this site are the rights of their respective owners

The CPU Site is not responsible for any kind of damage, based on information from this article

© 1998 CPU SITE futuretech@cpusite.demon.nl This site is sponsored by Demon Internet