Showdown at 133 MHz FSB – Part 2, The Real McCoy – THG.RU

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!

The RDRAM-Story, Continued

The PCXXX-specs are a very misleading designation, because the number behind the ‘PC’ is only reporting the memory clock. This makes sense as long as it is used for the same memory type, but it’s completely ridiculous if it’s used to compare two memory designs that are as different as SDRAM and RDRAM. SDRAM has a parallel interface that is 64 bit wide. Clocking SDRAM at 100 MHz leads to 64 bit or 8 byte * 100 MHz = 800 MB/s for PC100 SDRAM. RDRAM is a serial interface that’s only 16 bit wide and it’s taking advantage of ‘double-pumping’, so it can transport data at the rising and falling edge of the clock. Clocking RDRAM at 100 MHz would results in 16 bit or 2 byte * 2 * 100 MHz = 400 MB/s and it would be called PC200 RDRAM. You can see that RDRAM would require ‘PC400’-spec to offer the same bandwidth as SDRAM at PC100 spec. Thus PC800 RDRAM is only offering double the bandwidth of PC100 SDRAM, although it sounds as if it was eight times as fast.

The second problem of RDRAM is its latency. Now what the hell is that? I first decided to go on about data alignment issues and detailed comparisons between the SDRAM and the RDRAM interface, but I finally found a much better way to explain it. As already said, the SDRAM is a parallel interface, while RDRAM is serial. This is why SDRAM doesn’t need to be clocked as fast as RDRAM to reach the same bandwidth. At the same time the parallel interface has another big advantage. Just think of 20 motorcyclists on their bikes driving next to each other down a street and waiting for directions. Suddenly they are getting the command ‘turn right now!’. Because of the small size of a motorbike and because all of the guys are next to each other, they can turn into this street very fast, almost instantly. This was the example for SDRAM; it ‘turns off into little streets very fast’. Now picture a big bus with 20 people in it. The bus driver is also waiting for directions. When he gets the ‘turn right now’-command he will have to break and if it’s a very tight bend he even might have to reverse to get the big bus around the corner. This was RDRAM, it ‘turns around corners rather slowly’.

To get this comparison even closer to reality, you’ve got to picture the bus going at double the speed of the motorcyclists. Alternatively you could think of 100 fighter planes from WW2 and a Jumbo Jet. The fighter planes are not as fast, but they can bank a whole lot faster than the Jumbo. The only significant difference between these examples and RDRAM is the fact that RDRAM banks faster if it goes faster, not the other way around.

This is the big trouble of RDRAM. It has very high data bandwidth, but high latencies as well. If software requires data that’s spread all over memory, it has a serious performance problem. If data comes out of memory in big chunks, RDRAM can easily smoke SDRAM.

PC800, PC700 and PC600 RDRAM

PC800 is currently the fastest RDRAM available, but it is very difficult to produce it. The so-called ‘yields’ of PC800 RDRAM are very low and thus it’s very expensive and in low supply. PC800 RDRAM runs at a clock of 400 MHz and it uses a DDR-way of transporting the data. Thus the 16 bit or 2 byte data are transported at 800 MHz, explaining the ‘800’ in PC800. The data bandwidth of PC800 RDRAM is 1.6 GB/s and thus double of PC100 SDRAM or 1.6 times as much as PC133 SDRAM.

The story is a bit different with PC700 and PC600. Currently the RDRAM of choice is PC700 RDRAM. It was actually invented very late last year, when it turned out that the yields of PC800 were rather low. Intel’s 840 chipset is not able to use it; plugging PC700 into an i840-platform will result in the RDRAM running at PC600-spec, but i820 is able to make proper usage of PC700 RDRAM. This RDRAM-type runs at 356 MHz clock, offering around 1.4 GB/s data bandwidth.

PC600 is the cheapest and slowest RDRAM available for PCs right now, but it is still not coming at an attractive price point compared to SDRAM. If you think that PC600 stands for 300 MHz you’re dead wrong. PC600 RDRAM runs at 266 MHz, resulting in a data bandwidth of 1.066 GB/s, which is identical to the data bandwidth of PC133 SDRAM.

Intel 820 Chipset

I’ve written so much about i820 and ‘Caminogate’ that I will try to keep myself short this time. Of course i820 is using Intel’s new ‘hub architecture’, it supports AGP4x and ATA66. The i820-chipset was designed to use RDRAM in the first place and later Intel added a peculiar kind of SDRAM-support. i820 sports one Rambus-channel, which means that it offers one 16 bit wide RDRAM-interface. Thus the highest memory bandwidth would be 1.6 GB/s in case of PC800 RDRAM or less in case that PC700 or only PC600 are used.

Alternatively i820 can be equipped with a ‘MTH’-chip that offers PC100 SDRAM-support to save costs. Unfortunately the MTH re-translates the Rambus-protocol back to SDRAM-protocol, so that a system with this configuration can neither take advantage of the high data bandwidth of RDRAM (because it’s not used), nor of the lower latency of SDRAM, because the SDRAM doesn’t talk to the CPU directly, but ‘through’ the Rambus-protocol, which is already loaded with high latency. The only interesting thing about the MTH is the fact that it always runs at the 400 MHz that PC800 RDRAM would run at. Therefore the SDRAM ‘behind’ the MTH receives the commands at a faster rate than PC700 or PC600 RDRAM would. You will see why this is an important consideration when you look at the benchmark results below.

Intel 840 Chipset

The i840 was designed for workstations and small to midsize servers. The most significant difference to i820 are the two Rambus-channels that i840 is hosting. Both channels need to be equipped with the same amount of RDRAM and so i840 can take advantage of ‘interleaving’. This interleaving of the two channels can almost halve RDRAM’s latency and at the same time the two channels can together supply double the amount of data. Therefore i840 is less sensitive to software that jumps all around memory (e.g. database software) due to the reduced latency and it can deal with much larger memory streams. With PC800 RDRAM i840 offers a data bandwidth of up to 3.2 GB/s, with PC600 RDRAM it’s 2.1 GB/s, which still beats i820 with PC800 RDRAM. i840 does not support the PC700-spec. It’s either fish or flesh, either PC800 or PC600.

Intel’s 840 is from the same generation as i820. Thus it also supports SDRAM as long as at least two MTH-chips are used. It also supports AGP4x, ATA66 and it uses the Intel hub architecture.

Intel 815 ‘Solano’ Chipset

Last year Intel was still despising PC133 as an unreliable solution, but at this time Intel didn’t know how much everybody would hate i820 and RDRAM. Finally Intel came around and changed the specs of the upcoming successor of i810. ‘Solano’ was supposed to be another low-end chipset with integrated graphics, but today it might be the most interesting product that Intel will release this year.

Solano or ‘i815’ will have an integrated graphics adapter, but this baby can be turned off and you can use your own fast AGP-graphics card instead. i815 will also use PC133 SDRAM instead of the dreaded RDRAM and it will also get the new ‘hub architecture’ design.

Maybe that doesn’t really excite you yet, but I bet it will once you’ve spotted the results we scored with BX. After all we should have to see i815 as the real successor of 440BX. The most important advantage that Solano has over BX will be AGP4x support, ATA66-support, official PC133 SDRAM support, AGP-bus running at 66 MHz and thus in spec and the hub architecture. There is no obvious disadvantage over BX, unless you see the integrated graphics adapter as some kind of nuisance. After all it can be switched off though. Anyway, there’s no reason why Solano shouldn’t score at least as well as BX, as a matter of fact it should score even better. So if BX133 looks good in our comparison here, Solano should look even better.

Test Setup

Office Application Benchmark BAPCo Sysmark 2000

I guess the biggest surprise hidden in those results is obviously the highest score achieved by the platform with the 440BX-chipset. The oldest chipset with the ‘old’ SDRAM scores even better than Intel’s great 840-chipset. The low latency of SDRAM plus the low overhead of BX produces better results than i840 with 3.2 GB/s data bandwidth, but high RDRAM-latency. VIA’s Apollo Pro 133A is far behind BX, but it’s on par with i820 using PC800 RDRAM. Once i820 is using PC700, PC600 or SDRAM with the MTH it is the slowest chipset in the comparison!!

Office Applications in Detail

Bryce 4 is a 3D modeling and rendering software. You can see that i840 scores even less than Apollo Pro 133A and way worse than BX133, because it depends much more on low latency supplied by SDRAM than the high memory bandwidth supplied by RDRAM. This is the first benchmark in this test that does NOT like RDRAM.

Office Applications in Detail

Corel Draw doesn’t make that much of a difference between the different configurations, but it preferres i840 with its high bandwidth and reasonable latency and 440BX with its very low latency, but low bandwidth. In this comparison VIA’s Apollo Pro 133 A does not look exactly great. It scores worse than anything else.

Elastic Reality is a video editing software that seems to like high memory bandwidth, which is why i840 is the clear winner. However, i840 with PC600 and i820 with any kind of RDRAM are scoring just as good or bad as 440BX or Apollo Pro 133A. Therefore the memory latency plays an important role too. The RDRAM-solutions with single Rambus-channel as in i820 are clearly losing out against PC133 SDRAM here, but i840 with PC800 RDRAM is just about saving RDRAM’s honor. This is also the first benchmark where i820 with PC100 SDRAM and the MTH is scoring significantly better than i820 with PC600 RDRAM. This cannot be explained by the memory bandwidth, because those two configurations are identical in this respect. However, MTH plus PC100 SDRAM runs at a 400 MHz Rambus-channel and has therefore a lower latency than the PC600 RDRAM running at a 266 MHz Rambus-channel. It is rather obvious that RDRAM is not looking exactly superior here.

Office Applications in Detail, Continued

Of all Office application benchmarks this might be the one Intel and Rambus likes best. The high scores of i840 as well as i820 with PC800 RDRAM show how well RDRAM can do. However, it takes PC800 RDRAM and no less. PC700 and PC600 RDRAM score worse than i820 with PC100 SDRAM and the MTH. Apollo Pro 133A does not look great in this test and even BX133 is not scoring well either.

This benchmark shows how different a software application benchmark can look. Intel used Naturally Speaking in its own ‘Application Launcher’ benchmark, where it depended heavily on memory bandwidth. The results here show the opposite. BX133 is leading the pack and then VIA’s Apollo Pro 133A comes second. Intel’s 840 chipset looks pretty pathetic and i820 ‘Caminogate’ is even further behind. What conclusion does that lead to? Naturally Speaking does not like RDRAM?

Netscape Communicator again prefers BX133, something we will see a lot more often as we look through the results. i840 is at least scoring second, but i820 with only one Rambus-channel looks pretty bad indeed. VIA’s Apollo Pro 133A is clearly better.

Office Applications in Detail, Continued

Here we have it! Intel’s and Rambus’ nightmare comes through with this database benchmark. Paradox is jumping all over memory and the high latency of RDRAM makes it look really bad. VIA’s Apollo Pro 133A comes in second, clearly in front of Intel’s great i840 chipset. i820 with PC100 SDRAM and the MTH scores even better than with PC700 RDRAM. This should really make us wonder …

We are starting to get used to the unbelievable, BX133 is again the leader of the pack, i840 with the expensive PC800 RDRAM is just about making second, and Apollo Pro 133A beats i820 rather badly. Again the MTH-solution of i820 is ahead of PC700 RDRAM.

Power Point seems to use big chunks of data, which is why BX133 and Apollo Pro 133A don’t look too good in this test. Again i820 with the MTH and PC100 SDRAM is on par with the expensive RDRAM solutions.

Office Applications in Detail, Continued

This video editing software shows the dual Rambus-channel solution i840 in a pretty good light, but BX133 scores just as high as i840. VIA’s Apollo Pro 133A does not look exactly great in this test and neither does i820.

Windows Media Encoder favors i840 and BX133. Apollo Pro 133A may lag behind, but i820 scores ridiculously bad. Where is Camino’s greatness now, dear Intel?

Good old Word favors i840, BX133 scores second and Apollo Pro 133A looks rather bad. Still i820 with the MTH comes dangerously close to i820 with ‘fast’ RDRAM once more.

3D Gaming Benchmarks

I could hardly believe my eyes when I scored the result on the BX-platform! I had to re-run it several times until I grasped the fact that BX133 is indeed destroying all the younger competition. If you despise those results of an overclocked old chipset, please look at the scores of Apollo Pro 133A and compare it to i820 with PC700 RDRAM. Can you see that both Rambus-platforms score worse than the VIA chipset unless PC800 RDRAM is used? You are aware of the fact that most i820-RDRAM-systems ship with PC700 RDRAM. Those systems are badly destroyed by the much cheaper solution from VIA.

Again BX133 is killing all the younger and more expensive competition. Intel’s 840 with PC800 RDRAM comes in second, but again i820 with PC700 or PC600 RDRAM scores worse than VIA’s Apollo Pro 133A.

3D Gaming Benchmarks, Continued

Expendable shows similar results as the two Id-games above, but not quite as extreme. Again BX133 leads the pack, but the RDRAM-platforms don’t look quite as bad. Apollo Pro 133A looks pretty good too, especially if you consider its price.

Unreal Tournament’s UTBench is our final gaming benchmark that shows the superiority of the old 440BX chipset at 133 MHz FSB. Intel’s 840 chipset comes in second, but then there is VIA’s Apollo Pro 133A to show that i820 is only looking good with PC800 RDRAM.

Professional OpenGL Software

You can see that 77.7 fps seem to be GeForce’s fill rate limit in this test, which is why all the top players score the same. Again BX133 is amongst the leaders, showing how well an SDRAM-solution can compete against the latest RDRAM-chipsets. VIA’s Apollo Pro 133A does not look too great in this comparison.

These results give you quite a bit to think. BX133 scores best, which proves that high data bandwidth cannot be the explanation for the scores. VIA’s Apollo Pro 133A is again looking rather slow, while the Rambus-solutions are scoring very well.

Professional OpenGL Software, Continued

The story becomes a bit less difficult with Data Explorer. This benchmark is obviously favoring high data bandwidth, which is why RDRAM-solutions score very well. Still low memory latency must be a big issue too, because otherwise BX133 with its little 1 GB/s data bandwidth could hardly have made it second. Again VIA’s Apollo Pro 133A scores rather badly, only explainable by its high chipset latency.

The very last benchmark is the only one that takes real advantage of AGP4x. In almost all other tests BX133 was amongst the first three, but this Pro Engineer simulation is almost putting BX onto the last place. The ProCDRS benchmark needs AGP4X really badly, because it is using a huge amount of triangles. AGP4X needs high memory bandwidth to work properly, which is why i840 scores best. Again VIA’s Apollo Pro 133A cannot really impress anyone, but for once it scores better than the good old BX-chipset.

Conclusions

If we summarize the results we get to the following conclusions:

• 440BX at 133 MHz FSB is not a fully valid solution because BX and the AGP are running beyond specifications. At the same time ‘BX133’ is scoring best in consumer and office applications. Therefore I recommend conservative people to stay away from this combination, but whoever is not afraid of overclocking will find that BX at 133 is not only one of the cheapest, but certainly the fastest solution for Intel’s Coppermine-processors.
• VIA’s Apollo Pro 133A is currently the most sensible chipset for the average Coppermine-user. Systems with this chipset come at a very attractive price point, since motherboards and PC133 SDRAM are cheap. People who are using professional OpenGL-software should stay away from it though.
• Intel’s 820 chipset is in my eyes simply obsolete. Why would anyone buy a system with it? The performance is rather bad and the price is high. By the time when Intel releases i815 or ‘Solano’, i820 will simply disappear. Until then I would advise everybody to stay away from it, especially in combination with PC700 RDRAM.
• The Intel 840 chipset has its place in the workstation area, where it performs very well. I am not that impressed with it anymore though, since I had to see how badly it got beaten by the old 440BX-chipset in a large number of benchmarks.

I know that the inclusion of an overclocked 440BX-platform will upset quite a few of you, but it will delight a lot of others at the same time. Even though this old chipset was running beyond spec, its exceptionally good results should make us wonder what is going on in the platform market right now. Can we really accept that the latest Intel-platforms with the super-expensive RDRAM are not even able to beat the ‘tuned’ version of its two-year-old predecessor? Shall we sit idly, nod and go buy those platforms without even getting any real benefit? For most computer users the new chipsets i820 and i840 present the biggest fraud in the IT-industry that I have personally ever encountered. RDRAM seems nothing but a very bad tasting joke, unless you really use workstation type OpenGL-software. Maybe we should wonder why AMD never jumped on the RDRAM-bandwagon. Maybe AMD was wiser to stay away from it and count on the upcoming DDR-SDRAM memory solution. Could it be that Intel is only trying to get its return of investment out of Rambus? I really wonder what is going on here, but I know for a fact that I will stick to the BX-platform in my personal systems or that I will switch to an AMD Athlon-platform. I am simply disappointed and can only shake my head.

VIA was wise to take advantage of Intel’s i820/i840 RDRAM disaster. The Apollo Pro 133A is looking very good against i820, and it even beats it as soon as an i820-system is equipped with the common PC700 RDRAM. However, there’s huge room for improvement. Intel’s old 440BX-chipset at 133 MHz FSB is scoring so much better than the Apollo Pro 133A that you wonder if VIA will ever be able to design a fast performing chipset. It also shows how much room for improvement is left for Athlon chipsets. If the KX133 had an SDRAM-performance close to BX, Athlon could run a whole lot faster.

Intel will release i815 or ‘Solano’ sometime in June 2000. If Solano isn’t suffering an artificial slow-down, it should perform even better than the overclocked BX-chipset. In this case all the RDRAM-chipsets would look really bad and nobody will find a reason to buy i820-systems anymore. By that time VIA will also have trouble selling Apollo Pro 133A chips.