Preview of VIA’s upcoming Apollo Pro+ 133 Chipset — THG.RU

Camino is slowly on its way …

It’s time for a new Slot1-platform. Intel’s BX-chipset is out for about a year now, and whilst nobody is really complaining about its performance, the technology has to step ahead to create new business in the chipset and motherboard area. At the beginning of 1999 we still expected `Camino’, Intel’s upcoming 820-chipset, to be released by June of this year, but due to the exorbitant price of the memory type that’s needed by Camino, Intel delayed the launch to the end of September. Camino will provide a new CPU `front side bus’-clock of 133 MHz and it will use the new and proprietary RDRAM or `direct Rambus RAM’. RDRAM is using a completely different memory interface than EDO or SDRAM and it’s supposed to offer a memory bandwidth of up to 1.6 GB/s. It has to be plugged into the new `RIMM’-sockets and due to the fact that it’s no open technology, every memory manufacturer has to pay license fees to Rambus for each sold RDRAM RIMM. Unfortunately RDRAM is currently several times as expensive as SDRAM and it is rather questionable if it will offer us a substantial performance gain. However, since Camino will also bring us AGP4x, offering an AGP-peak-bandwidth of about 1 GB/s, more memory bandwidth is needed to feed AGP 4x properly. The current peak-bandwidth of PC100 SDRAM is just about 800 MB/s, which is even less than the AGP 4x-bandwidth, so that AGP 4x wouldn’t make much sense running on a platform based on the current memory technology. Still it’s questionable if it will indeed take RDRAM to satisfy the requirements of new software and future AGP 4x hard and software. Camino is also supposed to support SDRAM, but it will take a special converter riser-card to translate the serial Rambus memory interface into the parallel SDRAM memory interface, let alone the latency issues it would cause. All in all, Camino doesn’t look too promising right now, since it will increase system costs one way or another, if it should be a very high price for RDRAM or a still respectable price for a riser card that will only bring you very moderate performance.

Is there any reason to get all excited about 133 MHz FSB, 1.6 GB/s memory bandwidth and AGP 4x?

The short answer is `NO’. Before I get to VIA’s solution, I’d still like to remind you of the time when the BX-chipset was released last year and we had the chance to take advantage of the new 100 MHz FSB and memory clock. The performance gain we achieved by switching from 66 to 100 MHz was between 0-5 % in business applications and 3-10% in 3D-games. The latter was mainly due to the reduced arbitration problems of the AGP. There was also not much more than a 0-10% gain in 3D-game performance from AGP 1x to AGP 2x. Now we should expect even less than these performance gains from Camino’s RDRAM architecture and AGP 4x, because the increase in CPU memory interface bandwidth is only 33%, as opposed to 50% when we switched from 66 to 100 MHz FSB, and unless games are using a huge amount of either triangles or large textures, the gain through AGP 4x will also be less than substantial. The main reason why we will try to get platforms that offer 133 MHz FSB will be the fact that future Pentium III processors will be based on this FSB clock. This means that there won’t be a PIII 600 that runs at 6 x 100 MHz, and the next step will be 666 MHz and then 733 MHz, all only achievable with a FSB of 133 MHz. Now some of you may ask «why not overclocking my BX-platform to 133 MHz?» This would certainly be an option, since BX runs fine at 133 MHz FSB as long as you are using SDRAM that can cope with this speed, but what you forget is that BX can only run the AGP at either the FSB clock or at 2/3 of the FSB clock, which makes 88 MHz. An AGP running at 88 MHz is way out of spec and it will cause crashes as soon as your 3D-game is heavily using the AGP for textures or even only triangles.

VIA’s Apollo Pro 133 offers a platform for 133 MHz FSB at much lower costs than Camino

VIA designed the Apollo Pro 133 chipset to offer an alternative to BX as well as Camino. Instead of depending on the expensive RDRAM it can use PC133 SDRAM and later PC133 DDR-SDRAM as well as virtual channel RAM. Apollo Pro 133 will therefore offer you a platform to run a Pentium III (or for all overclockers of course also a Pentium II) at 133 MHz FSB safely, it runs the AGP within spec at 66 MHz = Ѕ the FSB-clock, and you can take advantage of PC133 SDRAM, which is available right now and not much more expensive than PC100 SDRAM. The memory peak bandwidth of PC133 SDRAM is 1.066 GB/s and thus less than the bandwidth offered by RDRAM, but it should still be an improve over PC100 and it will be a much cheaper platform for 133 MHz FSB-processors.

Enough with the theory, give us the benchmarking facts! — The Test Setup

Now this is all pure theory, it’s still necessary to put this new PC133-platform to the test. We should expect at least some performance gain over BX, especially with memory intensive applications. I’ve tested VIA’s PC133 vs. BX with the following system:

This is a picture of the EpoX PC133 motherboard:

Enough with the theory, give us the benchmarking facts! — The Results

I used a small set of benchmarks, but it should cover pretty much each important area. You know that I don’t believe in synthetic benchmarks that don’t give you the slightest idea what real world performance gain you can really expect. Thus I refrained from some beautiful but meaningless memory benchmarks or 3D-benchmarks. I used Quake 2 as well, but due to its heavy use of AGP with TNT2 it crashed in the BX system at 133 MHz FSB and 88 MHz AGP clock. So I took those results out. The idea was to run BX at 600/100 MHz, at 600/133 MHz and VIA PC133 at 600/133 MHz as well as 600/100 MHz. Unfortunately the EpoX P2-133A turns off the L2-cache of PIII at 600/100 MHz with its latest BIOS revision, which is why those results were removed as well. Please note that the only difference between a PIII-system clocked at 600/100 MHz vs. 600/133 MHz is the memory bandwidth. The L2-cache of PIII is in both cases running at half the CPU clock and thus at 300 MHz, the PCI-bus is in all cases running at 33 MHz, because even BX can divide the FSB-clock by four to get to 33 MHz. Therefore every difference in the benchmark results can only be explained by a difference in memory bandwidth.

Winstone was used to give you some idea about the office application performance, the Expendable Demo running in timedemo-mode was used to show the Direct3D-gaming performance and last but not least I used the Naturally Speaking 3.52-benchmark to test the memory performance, since this benchmark of Dragon’s well known voice recognition software does hardly do anything else than stress the main memory really hard, whilst hardly requiring any graphics or AGP performance at all.

As you can see, the results are pretty sobering. It’s an old issue that VIA’s AGP-performance lacks behind the AGP-performance of Intel’s BX chipset, which is why VIA’s PC133 lacks some 9% behind BX in Expendable-performance. The same is also valid for VIA’s MVP3-chipset compared to Ali’s Aladdin V chipset. The office performance of the PC133-platform is indeed slightly better than BX at 100 MHz FSB and on par with BX at 133 MHz FSB. The most disappointing result is achieved in the memory stressing Naturally Speaking-benchmark though. Even BX at 100 MHz FSB is already faster than VIA’s PC133-platform, but BX at 133 MHz is way faster. The BX-results are also showing how much Naturally Speaking stresses main memory. An increase in memory bandwidth of 33% leads to a surprising performance increase of 9%!! There are hardly any other benchmarks, which are depending on main memory as much.

Summary

Today at Computex I had the chance to speak to Wen-chi Chen, the President and CEO of VIA. He was not aware of the rather low performance of the Apollo Pro+ 133 chipset, but so far VIA has only used some synthetic memory and CPU-benchmarks. Many motherboard manufacturers including EpoX could also not give me any serious input about their performance expectations of VIA’s PC133-chipset. However, FIC and Asus were frankly telling me that they also found the VIA PC133-solution lacking behind Intel’s current BX-platform. Enhanced Memory Systems asked me to test their PC133 HSDRAM on an overclocked BX-platform rather than on a PC133-motherboard, because they were disappointed with the results of Apollo Pro+ 133 too. I hope that VIA can put some more work into the chipset as well as into the BIOS of the PC133 motherboards, to offer us a better performance than BX at 100 MHz FSB. I really like the idea behind PC133 and it’s really about time that someone provides a competitive chipset to the almost almighty Intel products. What other hope do we have? Well, there’s still Ali and SiS, but there’s also AMD with K7 and the corresponding K7-chipset. One of the four could still prove that performance hungry people are not forced to wait for Intel’s Camino and the expensive RDRAM. Let’s see who will succeed.

Quick Chart of the Pro’s and Con’s of VIA’s Apollo Pro+ 133