Showdown at 133 MHz FSB - Part 2, The Real McCoy

Introduction

Only a few days ago we published our Performance Showdown at 133 MHz FSB , where we tried to find the best mainstream platform for Intel's fancy Pentium III/Coppermine processor. We only looked at Intel's 820 chipset and VIA's Apollo Pro 133A, and the VIA-chipset came out as the clear price/performance winner. This review will be a lot different.

Intel's 440BX-Chipset At 133 MHz FSB

First of all I finally decided to include the good old and still very beloved 440BX-chipset from Intel, which is not certified for 133 MHz FSB. Still several people sent me emails claiming that an overclocked BX at 133 MHz FSB is a rather feasible solution. I tried it out and realized that I had to agree indeed.

RDRAM-platforms ...Why Not Testing PC700 And PC600 RDRAM As Well?

The next thing that all we reviewers tended to forget is an important issue with RDRAM. While all reviewers, including myself, used to do Intel the favor of using nothing else than PC800 RDRAM in our platform tests so far, the reality out in the computer stores looks a whole lot different. Due to serious problems in producing PC800 RDRAM, most RDRAM-systems that are shipping today are using PC700 RDRAM. We also included PC600 RDRAM in this test suite, because it is available, it's the cheapest RDRAM and it still claims superiority over SDRAM.

Intel 840 - The RDRAM King

Finally I decided to also include Intel's 840 chipset into this overview. Platforms with this chipset tend to be even more expensive than i820-platforms, but it offers by far the best performance you can get from an RDRAM-solution right now.

Considerations

This article will be filled with a huge amount of information and test data. To make this data easy enough to digest and understand, I will have to give a whole lot of explanations about the different platforms. If you want to make sure that you will understand the significance of the data I am going to present, I would suggest you read through those explanations beforehand.

Intel 440BX - Oldie But Goldie

It's now almost 2 years ago that Intel launched its first chipset for the 100 MHz front side bus (FSB), the 440BX. This chipset offered probably the most successful x86-platform that ever existed, and many people are still happily using it today, including myself. 440BX was designed for Intel's Slot1-processors at 100 MHz FSB and it is utilizing SDRAM with the PC100-spec, running at 100 MHz as well. Intel chipsets have a long history of being the fastest x86-chipsets on the market and from the day of BX's release until today there has never been any better or faster 100 MHz-chipset for the mainstream market.

As overclocking became more and more popular and after PC133 SDRAM for a memory clock of 133 MHz became available, people started running 440BX beyond spec at 133 MHz FSB. This procedure has the following issues.

• The front side bus of the CPU is running 133 MHz instead of 100 MHz, which is no problem even for processors marked for 100 MHz FSB, but certainly not for the latest Coppermine-CPUs that are supposed to run at this clock.
• The 'North Bridge' of BX is also running at 133 MHz instead of 100 MHz and thus out of spec. However, this does hardly do any harm to the chipset, it doesn't even get significantly hotter.
• The PCI-bus can still run at 33 MHz clock, because BX is able to divide the FSB-clock by 4 and 133 / 4 = 33.
• The 'South Bridge' of BX is not touched by this procedure, because it is connected to the north bridge via the PCI-bus. As long as the PCI-bus runs in spec, the south bridge does that too, so that I/O-ports, the integrated IDE controller and all the other components hosted by the south bridge continue to work normally.
• The only real problem with a BX overclocked to 133 MHz FSB is the AGP. AGP is supposed to run at 66 MHz and BX is able to ensure that by dividing the 100 MHz FSB by 1.5. Unfortunately Intel decided against the inclusion of a divider of 2 as well, which is why the AGP is doomed to running at 133 / 1.5 = 88.8 MHz. Running the AGP 33% beyond spec can produce a fair amount of trouble with AGP-graphics cards. Some have no problem, but many will simply freeze the whole system as soon as you switch to a 3D-application that is using the AGP. Fortunately our reference graphics card with NVIDIA's GeForce256-chip is not troubled by 89 MHz AGP-clock at all, so that the testing with a BX-board at 133 MHZ FSB was a piece of cake. The system remained absolutely stable, and you can believe me that I mean what I say, different to so many other overclocking-horney people, who think that a system that needs rebooting only three times/day is already 'rock stable'.

Intel 440BX - Oldie But Goldie, Continued

440BX has not been designed with Intel's fancy new 'hub architecture', but it proves that it doesn't require it either. As already said, BX's north and south bridges are classically connected via the PCI-bus, so that I/O-data travels through the south bridge over the PCI-bus to the north bridge and then to the processor. The PCI-bus has a data bandwidth of up to 133 MB/s and so far there's hardly any IDE hard drive or other peripheral device that comes close enough to that right now. The 'hub architecture' of i810, i810E, i820 and i840 connects the 'MCH' (ex north bridge) and the 'ICH' (ex south bridge) with a special bus that has a bandwidth of 266 MB/s and the ICH is in charge of I/O, IDE and PCI. BX is also missing AGP4x, since it only comes with AGP2x and its IDE-interface only supports ATA33 or UDMA33.

It's difficult to give an estimation on how BX at 133 MHz FSB (from now on addressed with 'BX133') would perform against Intel's official 133 MHz FSB chipsets or against VIA's PC133 chipset.

• BX133 almost has to lose out with software that takes advantage of AGP4x, but there's pretty little around. We also shouldn't forget that BX133 is not exactly limited to AGP2x. The AGP running at 88 MHz instead of 66 MHz makes BX133 behave as if it had 'AGP2.7x'-spec.
• There should hardly be any doubt that BX was designed to get the most out of SDRAM, which is why it should perform very well versus VIA's Apollo Pro 133A. Intel is known to design chipsets that have very low data overheads (the time that it takes for data to travel through the chipset, e.g. from memory to the processor), while VIA has always lost against Intel when competing with the same kind of memory interface.
• It is really hard to say how BX133 will look in comparison to i820 or i840 when those two are using RDRAM, because this new and very expensive memory type offers a much higher bandwidth than SDRAM does. At the same time there is the well known latency-problem of RDRAM, and here BX133 really shines, because it is using SDRAM, which comes with a very low latency and it comes from Intel, which adds very low overhead as well. Let's see what the results will tell us.

VIA Apollo Pro 133A

This new Coppermine-chipset from VIA was already introduced by me in the previous performance showdown article part 1 . The Apollo Pro 133A scored very well versus Intel's 820-chipset, almost reaching the same performance in most benchmarks. Competing against 'Caminogate' however is not the same as competing against the once famous BX-chipset. Apollo Pro 133A will have to show if its PC133 SDRAM-interface is up to competing against the so far best SDRAM chipset ever designed. I will only give this new VIA chipset my final blessing if it can live up to BX. If VIA has still not achieved that, it should send its engineers back to the drawing board.

The RDRAM-Story

I am sorry to say it, but I feel a tad bit repulsed to write about RDRAM. I doubt that there has been any other x86-component, which was hyped or lied about more than Rambus and its 'great' RDRAM. Still it is important that at least somebody tries and tells the truth about it.

When you are reading reviews about systems using RDRAM you will be presented with test data using RDRAM at the so-called 'PC800' spec. 'PC800' sounds tremendously of course and it certainly scores points with its name already, which sounds 8 times as fast, important (or was it expensive?) as the well known PC100 SDRAM. Long live the marketing departments of Rambus and Intel and long live all the people who believe that 'PC800' has any deeper meaning!