New Hope For Athlon — The VIA Apollo KX133 Chipset — THG.RU

Introduction

AMD’s Athlon-processor is a good and very well performing CPU. That’s something we’ve learned in August 1999, when it was finally released and stole the performance crown from Intel. Many people were rather delighted that Intel finally got some really serious competition and so the Athlon made its claim to fame. Intel had quite a lot of trouble to keep up with this new AMD-processor, but finally the chip-giant was able to reclaim the lead with the rushed release of the Pentium 800, using the well-known ‘Coppermine’-core.

A Little Quiz — Why is Pentium III faster than Athlon?

The current status in the x86-processor scene shows Intel’s Pentium III 800 (or ‘CuMine800’) as the leader and AMD’s Athlon 800 as the number two, pretty close behind it. This ‘I am number one and you are number two’-stuff is only something for the mentally not too challenged people though, since the whole story is a lot more complicated than some pure benchmark numbers comparing Athlon and Coppermine are able to reveal. The real secret lies in the ‘platform-issue’.

The truth, if I may call it that, is that

1. Pentium III 800 processors are very hard to get. Intel is able to produce some, but certainly not many of them.
2. Pentium III 800 is currently only really fast on platforms that depend on the super-expensive RDRAM-memory. Using the i820 or better the i840 chipset with this RAM makes Coppermine really fly, but unfortunately there are very few people who are willing to pay the very high premium for this memory-type.
3. many Pentium III 800-systems that are on sale are actually using VIA’s Apollo Pro 133 chipset or i820 with the ‘MTH decelerator chip’. Both types are using the much more reasonably priced SDRAM. The Apollo Pro 133 uses PC133 SDRAM, which gives the user at least some of Coppermine’s performance, i820 plus the cursed MTH can only use PC100 SDRAM and slows even this memory down so badly, that the performance is rather mediocre.
4. i810e may be an alternative for some people in Intel’s management, but it’s certainly one of the funkiest ideas to place a high-tech Coppermine-processor in a low-end i810e-motherboard. Would those Intel-managers once look at benchmark data of Pentium III on i810e vs. Athlon on Irongate they’d drop this idea pretty fast.
5. other people are using the well-established, but 2-year old 440BX chipset. Those guys have mostly upgraded from a previous Pentium III-processor and are at least not wasting any money on memory or second-rate platforms. However, on BX-platforms Coppermine can only run at 100 MHz front side bus and it is also unable to take advantage of AGP4x. Thus the BX-solution besides VIA’s Apollo Pro 133 is certainly the Coppermine-platform with the best price/performance ratio, but it still can’t take full advantage of Intel’s latest and greatest processor.
6. AMD’s Athlon is currently still damned to run on platforms using the one-and-only Athlon chipset ‘Irongate’, also manufactured by AMD. This chipset was initially still good enough to compete against all Intel processors on any platform, but it was already pretty outdated when it came to market. Today it is certainly not quite up to the latest standards anymore and we should therefore not be surprised that even the fastest Athlon can’t beat Intel’s fastest Pentium III processor, as long as it’s slowed down by a second-rate chipset.

AMD’s Struggle to Keep the Industry Interested

The main reason why it took so long until another Athlon-chipset came out is very easy to name. The reason is called ‘AMD’. Before you get upset let me please explain why I say that:

• AMD may have a respectable market share in the x86-processor business and it’s certainly the number two after Intel, but how far behind Intel is AMD? It’s far enough behind that many companies can’t be bothered to put any high emphasis on developments for an AMD-processor. Last year we could see that it was even worse. Taiwanese motherboard-makers couldn’t be bothered to offer Athlon-platforms, even though AMD could have equipped each motherboard with a CPU. That was even at the time when AMD was the clear performance leader! AMD may like it or not, its name doesn’t pull enough weight in comparison to Intel’s.
• The organization and infrastructure of AMD is also not comparable to Intel’s mighty empire. Until only recently, Intel stood for extremely high reliability in the business world. A company that worked together with Intel could look forward to fat profits and smooth operations. That might have changed with the i820-disaster and Intel’s recent inability to ship desired product, but CEOs are still sleeping better (although maybe still not good enough) with Intel as a partner instead of AMD.
• AMD relied on 3^rd-party chipset makers to come out with a new Athlon-chipset. Would Intel ever do that? Intel produces its own chipsets and until the i820-farce it used to have the reputation as the producer of the most advanced x86-chipsets on the market. We know that AMD does not have the resources to design a new Athlon-chipset, but what AMD tells you is that ‘they don’t want it‘.

Now many of us are Athlon-fans, but we’re still not stupid. An Intel-processor on a well performing and reasonably priced platform could still be more attractive than the beloved and excellent Athlon-processor on an outdated platform. The ‘Irongate’ is an outdated platform all-right, but AMD is currently really lucky that Intel is unable to supply a reasonably priced platform that performs well. Intel’s i820 and i840 chipsets are expensive due to the RDRAM-issue and a Coppermine-processor in an i810e-board is such a ridiculous idea, that not even the most inexperienced customers would fall for it. Intel’s new weapon will be i815 though and Athlon should get a new and modern platform before this chipset comes out.

The Release of the Apollo KX133 — Better Late Than Never!

After endless delays, VIA has finally released the Apollo KX133 chipset. This chipset was initially supposed to be launched right after AMD released Athlon. Compared with Irongate the feature list is rather impressive.

• PC133 memory, offering 33% higher memory bandwidth
• AGP4x-support, to take advantage of the latest 3D-chips
• ATA66 to improve the performance of the latest hard drives
• Two USB-hubs, resulting in four USB-ports
• AC97 low-cost audio and MC-97 low-cost modem support
• 2 GB max memory support

So far so good. However, the bad Tom always finds something to complain about. First of all I am asking for SMP-support. The KX133 does still only support one processor, so dream on you Athlon-fans with desire for multi-processor systems. The next thing I need to criticize is AGP4x combined with PC133 memory bandwidth. KX133 offers a peak memory bandwidth of about 1 GB/s, AGP4x can transfer data from main memory to the video-memory at exactly the same bandwidth. This means that no 3D-card will ever be able to take full advantage of AGP4x on a KX133-platform, because the CPU or other DMA-using devices will always require some memory bandwidth at the same time. Again VIA puts us off for ‘later’. At some time in the bright future there will be a successor of KX133 with DDR-SDRAM support, offering a peak memory-bandwidth of 2 GB/s. Let’s hope VIA can pull that off as quickly as NVIDIA did it in case of their GeForce-cards. I am having my doubts though…

The Block Diagram of Apollo KX133

Let’s be honest, it looks like all other previous AGP-chipsets from VIA and most of Intel’s too. There’s a north and a south bridge, both are connected with the PCI bus, the memory and AGP are directly hooked up to the north bridge… We’ve seen it hundreds of times, each time in a different color though. The point that I am trying to make is that VIA doesn’t have a funky ‘hub-architecture’ such as Intel’s latest 810,820 and 840-chipsets.

The Apollo KX133 Reference Board

Let’s get to the hands-on part of this review now.

VIA shipped me a sweet little reference board that looks like anything but a hot-rod. The ‘VT5249B1’ comes in MicroATX form factor, which means that you almost need to use a magnifying glass to find it. Plug the Athlon-processor with a massive heat sink in it and you get a rather funky picture.

This board is not meant to be for commercial purposes, so it can afford to have only two PCI-slots and one AMR-slot. Of course it has an AGP-slot also, just in case you couldn’t find it on the picture.

Getting the board to work was surprisingly painless. I plugged in the CPU, my reference GeForce DDR-card, a Netgear 10/100Mbit NIC, Micron PC133 SDRAM and switched on the whole enchilada. The system run instantly, which was one of the rather rare occasions in my history of pre-release hardware tests. Besides the normal stuff, the BIOS-setup is offering some rather kewl settings, which are not documented though. I turned on 133 MHz memory-bus, CAS delay of ‘2’ and AGP4x and started Windows98 without any problems. Then I adjusted the strange ‘R’-settings in the BIOS-setup one-by-one, until the system ran at the highest performance.

Benchmark Setup

Resolution for all 3D-games:

640×480, 16-bit Color, 85 Hz Refresh Rate

Resolution for Sysmark2000 under Windows98:

1024×768, 16-bit Color, 85 Hz Refresh Rate

Resolution for Sysmark2000 under WindowsNT:

1280×1024, 32-bit Color, 85 Hz Refresh Rate

Resolution for SPECviewperf under both OS:

1280×1024, 32-bit Color, 85 Hz Refresh Rate

Benchmark Results

To get a reasonable idea of KX133’s performance I ran BAPCo’s new Sysmark2000 under NT as well as Windows98, added some 3D-games for the toy-OS (Win98) and SPECviewperf for NT. I also tried my luck with Windows2000 Professional (build 2195), but so far KX133 and Windows2000 don’t really like each other, mainly because there’s no VIA-drivers for this upcoming new OS. Windows2k crashed as soon as I tried to install the NVIDIA GeForce drivers.

Sysmark2000 Results under Windows 98

This new office application benchmark does not show much of a difference between ‘Irongate’ and KX133. Two points are certainly not a lot. The reason is quite simple though. Office applications don’t really depend much on memory performance and they certainly don’t care about AGP4x.

Sysmark2000 Results under Windows NT

The difference between Irongate and KX133 is even less than above. Office applications simply don’t care much about those great new features.

3D-Game Applications under Windows 98

You can see that 3D-gaming apps benefit from KX133 quite considerably. +7 fps in Quake3 Arena are rather respectable and get Athlon 800 close to the Coppermine800 scores.

3D-Game Performance In Depth

To find out what makes KX133 faster in 3D-games, I played a bit with the BIOS-setup. I thought it would be interesting to find out if the 133 MHz memory-bus or the AGP4x are responsible for the higher scores. I also wanted to see how much the impact of CAS-delay of 2 vs. a CAS-delay of 3 clocks would be.

These results are very interesting. The main secret behind the performance-advantage of KX133 vs. Irongate seems to be PC133. The scores are highly depending on the memory-bandwidth, as you can even see if you compare the 133MHz setting with CAS2 to the one with CAS3. KX133 scores almost identical to Irongate once it runs at 100MHz memory clock. AGP4x vs. AGP2x vs. AGP1x is hardly making any difference. So far about the good old AGP-hype.

SPECviewperf 6.1.1 under Windows NT

The workstation-benchmark SPECviewperf 6.1.1 depends on both, memory-bandwidth as well as AGP-bandwidth, Let’s have a look at the scores.

The results are very surprising indeed. KX133 scores even worse than Irongate in AWadvs-03 and ProCDRS-02 (the two only benchmarks that are using textures), but it scores significantly better than Irongate in the other three. Could it be that VIA has still a bit of work to do? KX133 should never score worse than Irongate.

Again I tried the different BIOS-settings to find out what’s responsible for KX133’s performance. I altered AGP4x/AGP2x, PC133/PC100 and CAS2/CAS3.

The most surprising results are the ones where KX133 uses only 100 MHz memory clock. Here the AWadvs-03 and ProCDRS-02 scores are way less than the Irongate scores, the other results are almost identical to Irongate’s, just as we would expect it. As already mentioned, AWadvs-03 and ProCDRS-02 are using textures, the others don’t. Could this be the reason for the bad performance of KX133? Could it be possible that NVIDIA’s GeForce-driver has a problem with KX133 or does GeForce’s NT driver dislike Athlon altogether? It was definitely something to look into, which is why I decided to do something rather unusual, I ran SPECviewperf under Windows98. I couldn’t have been more surprised about what I found.

SPECviewperf 6.1.1 under Windows 98

All over a sudden Athlon scores a whole lot better and KX133 is faster than Irongate, just as it should be. The strange thing is though, that we should expect an OpenGL-workstation benchmark to run better under NT than under Windows98. I doubt than anyone would use high-end OpenGL-software under Windows98. Puzzled I looked at the Pentium III (Coppermine) scores and started to smell something seriously fishy.

What Does Athlon’s OpenGL Performance under NT Have in Common with an Excellent Munster?

I can’t help it, but that really stinks! Coppermine scores higher under NT than under Win98, which is what everybody would expect. It also scores way higher than Athlon under NT, something nobody would expect. Running SPECviewperf under Windows98 (what no sane person would ever do) reveals that Athlon’s Win98-scores are not only higher than Coppermine’s Win98-scores, they are even higher than some of Coppermine’s NT-scores! Athlon’s Win98-scores prove that this processor can run OpenGL-software on GeForce very well, but under NT, where those scores are important, Athlon performs miserably. Could it be that somebody who doesn’t want Athlon to look good on workstations is trying to make sure that either GeForce’s (and thus Quadro’s) NT-driver runs really bad as soon as it detects an Athlon-processor, or is there some strange anti-Athlon-software in WindowsNT’s latest service pack? Many of you will know that Quadro, a NVIDIA GeForce-chip with 64 MB DDR-memory, is an extremely good high-end OpenGL-performer and will thus sell very well in the workstation area. Making sure that Athlon performs really badly with this graphics card almost means barring Athlon’s well-deserved way into the workstation-segment. The SPECviewperf-scores under Windows98 prove that there’s certainly no hardware problem between GeForce and Athlon. They also prove that Athlon is actually performing better than Coppermine in OpenGL. The fishy scores under NT raise some serious questions towards NVIDIA, Microsoft and also AMD, since nobody there seems to have noticed this issue early enough.

You see, this is AMD-bully No. 2,345,687. It’s exactly what I was talking about in my introduction. Intel seems to pull enough weight that a well performing AMD-product is shut up with some cheesy software. If AMD isn’t finally starting to do something about those kinds of issues, the days of this company will be counted.

SPECviewperf Scores In Depth

I decided to make a in depth evaluation of each SPECviewperf ‘viewset’ with the different BIOS-settings, just to see if I can find more meaningful results than what I found with the 3D-games.

The ‘Advanced Visualizer‘ doesn’t seem to care much about memory bandwidth at all, and it also doesn’t make any difference between AGP2x and AGP4x. However there is a significant impact as soon as the AGP-speed comes down to AGP1x. It’s also interesting to see that Irongate is not even up to competing against the worst scores of KX133.

‘Design Review‘ is certainly fussy about the memory-bandwidth, as you can see even when you compare the results of PC133 and CAS 2 vs. CAS 3. AGP-speed does only have a small influence once it becomes less than AGP2x. Again, Irongate is behind all of KX133’s scores.

SPECviewperf Scores In Depth, Continued

‘Data Explorer‘ has a similar behavior as ‘Design Review’, it cares a lot about memory throughput and rather little about AGP-speed. Irongate is again the slowest.

‘Lightscape‘, the by far toughest of the five ‘viewsets’, is again depending more on memory-bandwidth than on AGP-speed.

‘ProCDRS‘, the most well-known of them all, shows a different pattern. Memory bandwidth seems just as important as AGP-performance, which is why this is the only chart where results are decreasing continuously. AMD’s 750 ‘Irongate’ chipset performs worst in this test as well. I’d also like to point out that KX133 is no less than 56% faster than Irongate in this benchmark. ProCDRS shows the potential of KX133 extremely well.

Conclusion

The long awaited Apollo KX133 chipset from VIA Technologies is without the slightest doubt a good product. In all benchmarks it performed better than AMD’s aging 750 ‘Irongate’-chipset, beating it in some by more than 50%. It is the one thing that AMD really needed to keep Athlon competitive with Intel’s Coppermine. In my eyes VIA was able to keep its promises for the very first time. I remember numerous previous VIA-chipsets that always looked great on paper but never lived up to its specs. The KX133-reference board is also the first pre-release VIA-motherboard I ever received, that worked without a flaw in a very instant. If this should be a sign for the direction in which VIA is heading, then I want to send my deepest ‘kudos’ over to Taiwan, the former ‘Formosa’. At the same time I’d like to remind however, that there’s still a lot of expectations that want to be fulfilled soon. KX133 needs a higher memory bandwidth to accommodate AGP4x in a more efficient way and I wonder if I will live to see an SMP-chipset from VIA Technologies any time soon.

If you are an end-user you certainly want to know if I recommend motherboards with this chipset. My answer is simple. If you are planning to buy an Athlon platform you should not possibly miss it. If you already own an Athlon-system with AMD’s 750 chipset and if you mainly run office applications or web-software then I can’t see any reason why you would have to swap your current motherboard with a KX133-board. Office applications hardly benefit from higher memory-throughput, something that makes Intel-platforms with i820 or i840 chipsets completely unattractive to office users as well. Gamers have to decide if the 5-7% in gaming performance offered by KX133 are reason enough to say goodbye to their old Irongate-boards.

If I knew how long it will take VIA to offer a KX133 version that supports DDR-SDRAM, I would give the same advice I gave all the people interested in GeForce. Wait until the DDR-version is available.

Follow-up by reading the article ‘NVIDIA’s New NT-Drivers for Athlon on KX133‘.