The 150 MHz Project, Part 1 – THG.RU

Introduction

One of the most important topics beside the processor speed is the bus clock of a CPU. The clock speed at which the Front Side Bus (that’s how Intel calls it) and the system is working has reached 133 MHz today. Some time ago, Tom already dealt with this topic and proved impressively that rising the bus speed can lead to better performance numbers than just using higher clocked CPUs (The Bus Speed Guide). As a byproduct of processor development, L2 caches have been added to the CPUs to increase performance and to gain more independence from the main memory.

Bus Speed History

In the beginning of computer history, there was no difference between the internal and external clock speed. A 386 DX 25 worked at 25 MHz processor clock and was able to communicate at the same 25 MHz with its page mode DRAM. In the beginning of the 90s it became obvious that processor speeds would rise faster than clock speeds of the main memory (best example was the 486 DX 50) and upcoming synchronous bus systems. Since the performance losses at lower clocked memory were tolerable, Intel introduced a new series of processors: The 486 DX2. The “2” means that the processor is clocked twice as high as the memory/bus. A 486 DX2 – 66 worked at 33 MHz memory/bus and 66 MHz processor clock.

Thus it was possible to use less expensive memory chips. Thanks to the increasing use of L2 caches, performance drops caused by the lower memory clock were small. The main reason for this “clock splitting” was the introduction of the VESA Local Bus and the first PCI systems, offering much faster bus performance for graphics and I/O components. While PCI was specified for only 33 MHz, VLB was specified for 40 MHz. To obey it, processors were clocked at double (486 DX2-50, DX2-66) or triple bus speed (486 DX4-100).

With the introduction of the first Pentium computers, the internal and external clock speeds had been equalized for the last time: Pentium 60, Pentium 66. Following processors (Pentium 75, 90, 100) worked with a multiplier of x1.5 and different bus speeds (50, 60, 66 MHz).

Today’s computer architectures have been optimized, so that some computers do already work with multipliers of up to x8 (Pentium III 800 at 100 MHz bus or Celeron 533 at 66 MHz bus clock). However, performance increases cannot hold up to the clock speed rising as long as the bus speed is not pushed up as well.

Common Bus Speeds

What exactly does happen when we chose a higher clock speed? As you know there is the system clock speed and the processor clock, which can be “produced” by multiplying the system speed with numbers between 1.5 (2.0, 2.5, 3.0…) and 8.0. All other clock speeds depend on the system clock. The PCI bus always runs at 1/2, 1/3 or 1/4 of the external clock just like the AGP runs at 2/3 (100 MHz bus) or 1/2 (133 MHz bus) of the system clock. This table will give you a little more explanation:

¹ There was no AGP at that time.

² The K5-PR133 works at 100 MHz clock speed. The K5-PR166 as well, but with an internal multiplier of x1.75. You have to set X2.5 on the motherboard.

³ Athlon processors work at 200 MHz FSB and 100 MHz system clock speed. The memory can be clocked at 100 MHz (AMD750/Irongate) or at 133 MHz (VIA KX133).

The table shows all possible processor clockings which are officially specified. Rising the bus clock means overclocking motherboard, controllers and main memory. Of course it’s possible to rise the bus speed at all three speed grades, but remember that dependent clock speeds do rise as well.

PCI and AGP Clock at Higher Bus Speeds

Processor overclocking by chosing a higher multiplier can only be done with Athlon CPUs, using a special overclocking board. Intel processors have a fixed multiplier with no option to change it. The only solution for speed freaks and overclockers is of course a higher bus clock. Be careful when rising it! The table above showed the coherencies in standard systems. Never forget the other clock speeds when overclocking:

¹ This setting will most likely fail, since add-on cards and on-board components are overclocked just too much.

Chosing a much higher bus clock than standard will cause instabilities resulting in hang ups or data losses. Thus it’s very important to run a reasonable bus clock which most likely won’t cause trouble. In the past, 75 instead of 66 MHz have proven to be very reliable (that’s the best way to overclock a Celeron processor by the way). For 100 MHz FSB, 110 or 112 MHz seemed to be the best choice.

The Test System

We decided to overclock by 10-15% again, since this seems to be a slew which most components should tolerate. Our basic system is now running at 133 MHz FSB by default. Again there is the question whether to use a system based on i820 and RDRAM, or a platform using VIA’s Apollo Pro 133A chipset. In my eyes Rambus systems are not particularly suited for overclocking due to the memory issue. i820 can only be faster using PC800 RDRAM, whose actual pricing is as high as low end computers. People who want to spend more money on their memory than others would spend for a CPU usually don’t need to overclock their system.

VIA’s Apollo Pro 133A chipset is a platform which is affordable and which should be quite widespread very soon.

Requirements for Bus Speed Overclocking

Basically you only have to ensure that all components work properly at the higher clock speed. If you overclock a system by rising the processor core speed (by changing the multiplier), you can be sure that the processor will be responsible for all types of troubles. Of course this is quite different with bus speed overclocking: All related clocks and components will run faster as well: Chipset north bridge including memory controller, the main memory, PCI and AGP and the CPU Front Side Bus. Even if only one component fails you will have to spend some time to find out which one.

1. Good memory
   
   Keep some reserves. With 66 MHz systems it’s best to use PC100 memory or at least brand 10ns types. They should usually bear 75 MHz without any complaint. For 112 MHz, take at least 8 ns PC100 SDRAM. 7 ns are of course better, but those speed numbers are not as important as having memory from a reliable manufacturer like Micron, SEC, Toshiba, Infineon, Viking or others. The best is of course to use PC133 memory, but I don’t think that many of you will buy new memory only for overclocking. I wouldn’t do it either.
   
   For overclocking from 133 to 150 MHz, we used good quality SDRAM from
   
   Crucial/Micron, which easily withstands most hardcore memory timings.

2. Quality components
   
   Each component you want to overclock has to work reliably at the higher clock speed. Critical components are network cards or SCSI adapters as well as some graphic cards. Make sure that your components run properly at your bus speed setting. Don’t expect cheap products to run properly. It is likely they won’t cause problems, but of course nobody can guarantee this to you. The limit for timing-sensitive PCI cards is usually 36-38 MHz (or about 75 MHz for AGP graphic cards), this means that you should do intensive testings with your new bus speed:
   
   Does the SCSI card work properly: Can I burn CDs without errors? Does the SCSI scanner work? Are all drives working well?
   
   A graphic card may boot up and show the Windows desktop, but what about 3D graphics? Does everything run as expected? Is the picture corrupted?

The 150 MHz Project

Of course I used our “standard benchmarks” for the 150 MHz speed tests: SYSmark 2000 using Windows 98 SE and Windows NT 4.0 SP6a, Quake III Arena and Expendable. The SYSmark consists of two parts: Internet Content Creation and Office Productivity. This time I will give you those single results as well, since it makes it easier to show which applications will benefit from higher bus speeds.

Internet Content Creation uses Bryce 4, Elastic Reality 3.1, Photoshop 5.5, Premiere 5.1 and the Windows Media Encoder 4.0. To get the Office Productivity result, BAPCo makes use of Corel Draw 9, Excel 2000, Naturally Speaking 4.0, Netscape Communicator 4.61, Paradox 9, Powerpoint 2000 and Word 2000.

We were using a Pentium IIIEB 667 for these tests. Luckily this engineering sample allows this multiplier to be changed. We did benchmarks at 100 MHz FSB, clocking the CPU at 650, 700, 750 and 800 MHz. At 133 MHz FSB, we took 667, 733 and 800 MHz. I would have loved to give you the results at 825/150 MHz, but the CPU didn’t allow the SYSmark to complete. Therefore we can merely provide the game benchmarks, although I think they are quite impressive.

Applicaton Benchmarks Windows98 SE

Using Windows 98, a Pentium III clocked at 750/150 MHz is clearly faster than a Pentium III 800/100, but slower than the top model at 133 MHz FSB.

To get the performance of a Pentium III 750/100, you can also get a Pentium III 600EB and run it at 150 MHz bus speed (675 MHz).

The same scenario. You can see quite well that 100 MHz FSB is not suited for high end computers any more. Using a 133 MHz CPU will give you the performance of the next speed grade of the 100 MHz pendant.

Applicaton Benchmarks Windows NT4

These results prove again that a Pentium III should not be limited to 100 MHz FSB. The 800 MHz type is about 6% faster thanks to the faster bus clock.

Office Applications seem to benefit a bit more thanks to the 150 MHz bus speed. A Pentium III 675/150 is even faster than a Pentium III 750/100 in the Office Productivity part of the SYSmark. The same applies to Pentium III 733/133 and 800/100.

Game Benchmarks

Quake III obviously loves fast bus speeds! The Pentium III models running at 100 MHz FSB are the clear losers in this comparison. A Pentium III clocked at 750/150 is about 16% faster than at 750/100. Even a 667/133 is faster than the Pentium III 800/100. The beloved OpenGL 3D-shooter is obviously moving around large data amounts. Now imagine what DDR memory could do…!

Expendable makes use of Microsoft’s Direct3D interface. It cannot benefit from ultra high clock speeds just like Quake III, but still the performance gains are bigger than with standard Windows applications.

Conclusion

Both gaming benchmarks have particularly shown the importance of the FSB. It’s still as present and mighty as three years ago, when we were running 66 MHz systems at 75 or 83 MHz. Even today it’s something like an unwritten law: Using a higher FSB will rise performance about one performance step.

Using 140 instead of 133 or 112 instead of 100 MHz will give you more performance with little risk. 150 MHz can be achieved quite safe – provided that you are using quality components (especially motherboard and RAM).

You will have to spend some days on testing the new speed setting – in fact, the longer the better. If only one component fails (e.g. if writing CDs is not possible at your common speed), step back one FSB step.

Who Really Benefits from 150 MHz FSB?

Basically everybody, but gamers and users of high end software will benefit more. Overclocking always means a kind of adventure. Running hardware at 150 MHz is interesting for most users who do not necessarily need their computer, but want more performance – overclockers, gamers and performance freaks. All others should be very careful, particularly those who require a working computer.

What About 160 MHz or More?

The faster the bus speed, the higher the risk for your components. For these tests I used an Asus V6600 AGP. I originally wanted to do some benchmarks at 160 MHz as well, but the card did not want to work at 80 MHz AGP clock. The ATI Fury MAXX seemed to run fine, but after some minutes the system crashed. I’m still not sure what caused these crashes, if it was the graphic card or the motherboard.

I think this shows the limit of present computer technology. That’s also why the next memory generation will be able to transfer q2asdouble the amount of data – not by doubling the clock speed but by making use of both edges of a signal.

Follow-up by reading the article ‘The 150 MHz Project, Part 2‘.