The RDRAM Avenger – Intel’s i840 Chipset – THG.RU

Introduction

It is more than six weeks ago now. One day after Intel released its new Coppermine processor-core, exactly on October 26, 1999, I received an email from a member of Intel’s Customer Support who is working in the Server-Division. The mail was nice and polite, but stated that if I really would want to see Coppermine’s performance, I should run this new CPU on a platform with Intel’s i840-chipset, particularly on the ‘OR840’-motherboard from Intel. The sender stated that this is his personal opinion and not Intel’s (every Intel-employee is very careful about that) and noted that he will try and get an OR840 for himself as soon as it is available.

It was that day when I started to dig for one of those mystical ‘OR840’-boards, or at least another motherboard with this shiny new i840-chipset, that looked so much better than the ‘tainted’ i820.

The Intel 840 Chipset

The i840 is actually a ‘workstation’-chipset and the designated successor of the 440GX-chipset. Due to this, you won’t find i840 on mainstream motherboards right now and Intel doesn’t really like anyone to talk about i840, who does not address workstation users. However, looking at i840’s features and its differences to i820, the RDRAM-mainstream chipset, make i840 appear rather attractive to anyone who wants to use a Coppermine-processor.

On the first look i840 isn’t really that different to i820. Both chipsets share the RDRAM-support, AGP4x, the ‘accelerated hub architecture’ with the 82801 I/O Controller Hub (ICH), ATA66-support, dual CPU-support, and both chipsets sport the 133 MHz front side bus. So far both are the same; the difference are the added features of i840:

• Dual memory channels, operating in lock-step, provide up to 3.2 GB/s of memory bandwidth. Thus i840 has double the memory bandwidth of i820, which is a rather huge difference. The two memory channels can also reduce latency, so that i840’s memory interface is way better than i820’s. Additional to that, i840 supports up to 2 GB of memory, also double of what i820 can be equipped with.
• Using the optional 82803 RDRAM-based memory repeater hub (MRH-R) or 82804 SDRAM-based memory repeater hub (MRH-S) increases i840’s memory capacity even further, but the current cost of RDRAM and the performance loss of using SDRAM with either i820 or i840 doesn’t make this option look particularly attractive.
• i840 comes with a special prefetch cache, allowing more efficient data flow and improving system concurrency.
• The optional 82806 64-bit PCI Controller Hub (P64H) supports 64-bit PCI slots at speeds of either 33 or 66 MHz. The P64H connects directly to the MCH using Intel Accelerated Hub Architecture, providing a dedicated path for high performance I/O. This is a feature for workstations indeed, and you won’t find it on the less sophisticated i840-motherboards. Mainstream users won’t be able to take much advantage of 64-bit PCI-slots, since there’s hardly any 64-bit PCI-cards available in the first place and there would also hardly be any need for a 64-bit PCI sound card or network card for normal users.

In short, the main catch of i840 is it’s two RDRAM-channels versus only one RDRAM-channel of i820. This increases memory bandwidth and reduces memory latency, but you always have to plug two RDRAM-RIMMs into an i840-board. i840’s prefetch cache is another advantage over i820, and you can notice that in more than just workstation software, Quake3 likes to run on i840 as well.

The Intel OR840 Motherboard

The first motherboard available with the i840-chipset is clearly Intel’s OR840. It’s a dual-processor motherboard and comes with 10/100Mbit-LAN adapter onboard, of course using Intel’s 82995 chip. As you can see from the picture above, the OR840 is equipped with a AGP Pro50-slot. This slot hosts normal AGP-cards just as well, but it supplies additional power lines for professional graphic cards that need a lot of current. The picture shows the four RIMM-slots, always two acting as a pair for each RDRAM-channel. As already mentioned above, you need to install identical RIMMs for each channel, the board does not run with only one RIMM or two differently sized RIMMs. For my testing I plugged a terminator-module into the second ‘Slot1’ CPU-slot, the other one hosted a Pentium III 700 processor with Coppermine-core, overclocked to 750 MHz, or a Pentium III 733 processor, also using the Coppermine-chip inside. The rather massive voltage regulator area that you can see in the upper-middle part of the picture is obviously well designed for workstation use with components that need a lot of current, and there are some additional power-supply headers on the board as well, in case the normal ATX-power connector should not be able to supply all the current needed. As for all of the new Intel-motherboards, you won’t find any ISA-slots anymore. Another interesting point is obviously the price of OR840. I was informed that this dual processor workstation board comes for about $300 or less, which is not a bad price for all the features it adds to a $100 cheaper i820-motherboard for single-processors.

The testing with OR840 ran rather smoothly, the board never crashed or showed any other sign of instability. The adjustment of the processor-speed for unlocked processors is rather funky. You don’t change the multiplier in OR840’s BIOS-setup, but you need to run a little DOS-program called ‘speed.exe’ instead. I guess that this little side note is not too interesting to most of you, because all the Pentium III-processors that are officially available will be multiplier-locked and thus there’s no need for ‘speed.exe’.

There was one annoying thing with OR840 though, it wouldn’t like a lot of graphics cards. I used NVIDIA’s reference GeForce256-DDR card and it ran fine. So was the TNT2-Ultra reference card. However, no Diamond-card would run in OR840, neither the Viper V770, nor the Fire GL1. OR840 refused to even come to life with those cards. I am sure that a future BIOS-update will remedy this problem, but it kept me from running the SPECviewperf tests with the Fire GL1.

The Contestants

You need to compare, if you want a good idea of the performance of a product. I decided to compare i840 with its two brothers from Intel, the good old 440BX chipset that we all love and the i820 chipset, that many of us despise.

440BX

I guess that I don’t have to say much about BX. This 100 MHz FSB chipset, using PC100 SDRAM is now more than a year old and it still performs excellent. The motherboards available with BX are all mature and running rock-stable. It’s good for overclocking and hosts all a normal user needs. All in all the BX-chipset will stay for quite some time to come, because it’s far from outdated and it has a lot of fans and supporters. I used my favorite BX-motherboard for this test, the Asus P3B-F.

i820

I don’t really want to bore you, so I won’t go on too much about i820. There’s hardly any chipset that got so much press prior to its release. First known under the code name ‘Camino’, it was supposed to be the first chipset to bring us the blessings of RDRAM. Now we know that RDRAM is a serious political issue in itself already, because RAMBUS, its developer, receives royalties for each sold RDRAM-module. At the same time Intel holds a lot of shares of RAMBUS and is its only real promoter. RDRAM has got a very different interface to the memory-types we are used to so far. It uses a 16-bit wide serial bus that runs at 300-400 MHz, which enables a very high memory-bandwidth but at some penalty in terms of memory-latency. RDRAM is still super-expensive, costing about $1000 for 128 MB. The RDRAM-issue may be valid for i840 as well, but i820 had and has some more problems. Intel had to delay its release, because it turned out to be unfeasible to design an i820-motherboard that offers three RIMM-slots and make sure it’s stable at the same time. The crop of i820-board becoming available now has thus only got two RDRAM memory-slots. At the same time i820’s performance is not exactly impressive. There are several benchmarks where BX looks at least as good as, if not even better than i820. As a summary, an i820-system is expensive due to the RDRAM-price and it doesn’t even perform better than BX. However, i820 is THE mainstream chipset from Intel with support for the new Coppermine-Pentium III processors running with 133 MHz front side bus. This means that you either got to stick to 100 MHz FSB processors and use BX, or you have to go for VIA’s Apollo Pro 133+ chipset, which offers 133 MHz FSB with PC133 SDRAM. You can see that Intel hopes that it can push i820 through into the market, regardless if this product is good or bad and regardless if RDRAM is ridiculously expensive or not. However, there is one alternative, the i840-chipset. Let’s see if i840 is able to at least add above-average performance to the very high cost of memory, something that i820 was not able to achieve.

i820, Continued

My i820-testplaform was an Asus P3C-L motherboard.

This new i820-motherboard is the main reason why my review got delayed by about three days. Quite a while ago and still prior to i820’s final release we received i820-motherboards from Intel that came with the initial three RIMM-slots. For some reason the performance of those nowadays obsolete platforms was significantly better than the new crop of the official i820-boards with only 2 RIMM-slots. As a matter of fact the new i820-motherboards score worse than BX in several NT-benchmarks, as you will see later. This is not an issue with the Asus P3C alone, all the other i820-boards we tested scored just as bad. Unfortunately Intel was unable to give us one of their new boards with only 2 RIMM-slots, so that we couldn’t see if the performance of those boards dropped as well. My experience tells me however, that Intel-motherboards have never performed better than Taiwanese ones; usually the Intel-product was a bit slower. It would be very surprising if that should have changed with the release of i820.

The P3C-L that I used for testing sports some kind of ‘all-in-one’ PCB. You can see the free areas in the upper right corner of the picture. This is where Asus can put a SCSI-chip and/or an IEEE 1394 chip. One of the tiny chips in the front is the Intel 82559-network chip, which is responsible for the ‘L’ in the name of this Asus-board. You can also see an ISA-slot, which Asus made possible by using a PCI-ISA-bridge from ITE. The AGP-slot is an AGP Pro50 slot, the same as in case of the OR840. Next to it you can find the AMR (audio-modem-riser) slot.

Ultra-ATA

You are certainly aware of it, the ‘ICH’ = I/O controller hub of i820 and i840 (it’s the same for both) is capable of ‘UltraDMA66’ or ‘UDMA’ or ‘ATA66’, which stands for a maximum data transfer bandwidth of 66 MB/s between the system and the IDE-hard drive, as long as the hard drive supports it. The 440BX chipset is only able to do ‘ATA33’ = up to 33MB/s between system and IDE HDD.

Now which driver supports the new ATA66 under Windows98 and what do you have to use under WindowsNT even for ATA33? Intel has finally released a pretty swanky driver, which you can find here. This driver works for Windows98 as well as for WindowsNT, although the install-routine under NT has got a little bug that you will notice after installation, if you look into the ‘SCSI-Devices’ control tab. There you won’t find the new driver, although it’s running fine. Instead you’ll see the old standard-IDE driver marked with ‘not started’. Anyway, whoever is using IDE disks and Intel-chipsets should have a look at Intel’s new Ultra ATA driver.

Benchmark Expectations

What should we see when testing the different chipsets against each other? Well, I would like to see superior numbers of i840, showing its impressive 3.2 GB/s memory bandwidth, which is double the bandwidth of i820 and even quadruple the one of BX. BX should look bad in 3D-applications, since it’s the only one of the three without AGP4X-support. i820 should be somewhere in between i840 and BX, especially in memory-intensive applications. OK, can you agree with that?

Benchmark Setup

For Business Application Testing I used the well-known BAPCo Sysmark98 under Windows98 as well as WindowsNT and Intel’s Business Application Launcher 1.2 under NT only. For 3D gaming I ran the retail-version of Quake 3 Arena, the retail version of Descent 3, the demo version of Expendable and 3DMark2000 from MadOnion.com. Workstation-performance under NT was checked out with SPECviewperf 6.1.1 and ZD’s Highend Winstone98. Unfortunatlely Winstone2000 is not available yet and Winstone99 would not run AVS with the latest driver of GeForce.

Business Application Performance under Windows98

The world looks normal here. BX makes third and thus last, i820 is a bit faster and i840 puts itself quite well ahead of the rest with a comfortable lead of no less than 5 points.

Business Application Performance under WindowsNT

Here is what I couldn’t believe at first. i820 scores worst, quite a bit behind BX and far behind i840. I wonder what has happened to the latest i820-platforms, they score bad only under NT.

Business Application Performance under WindowsNT, Continued

We can see almost the same with Intel’s Business Application Launcher. i820 scores really bad in Naturally Speaking, which used to be one of its favorite benchmarks.

In Netshow Encoder the world looks normal again, i840 scores best, closely followed by i820 and BX comes last.

Photoshop doesn’t seem to care too much about the platform it runs on. All chipsets score pretty close to identical.

3D Gaming Performance Windows98

It doesn’t seem to be AGP4x that’s making the difference here, i840 smokes the rest and i820 can’t score any better than BX at 750/100 MHz, although it comes with AGP4x versus the AGP2x of BX. It’s got to be i840’s superior memory bandwidth that makes it score so well. I guess the i840 is the new top platform for Quake 3 Arena players.

The picture is pretty much the same. i820 only scores at 133 MHz FSB, otherwise it performs almost identical to 440BX. i840 is the new Coppermine platform of choice for gamers.

3D Gaming Performance Windows98, Continued

At last in Expendable i820 is able to regain some honor versus BX. i840 is still leading the pack though.

The results of 3DMark2000 are pretty much the same as what we’ve seen in Expendable. BX scores worst, but i820 is still behind i840.

Workstation Application Performance under WindowsNT

Now we know why i840 is a workstation chipset! The i840-scores are up to 40% higher than the scores of BX, still up to 8% higher than i820. It’s true, for a workstation user it may indeed be worth getting rid of his BX/GX-platform and investing in i840 and the expensive RDRAM.

These results remind us of the office application performance and the Sysmark98-scores. BX is as fast as, actually even a tad faster than i820. Why would anyone be interested in this chipset?

Summary

You’ve seen the results, so you might not even need my advice anymore. For anyone sensible looking for a new platform i820 should be pretty much out of the question.

This half-baked Intel-chipset doesn’t perform well enough and requires very expensive memory too. If you are on a budget go for BX, consider VIA’s PC133 SDRAM chipset or wait until the DDR-platforms become available.

Brent will post a first review of a DDR-platform soon. For the performance-hungry Coppermine-freaks there’s only one chipset, i840! Motherboards with this workstation chipset may be a bit more expensive than i820-platforms, but you get top performance and some other nice features as e.g. a dual-board for it.

Considering the high costs for RDRAM, the 100 bucks more for i840 over i820 are certainly well spent and this expensive memory is completely wasted should you plug it into an i820-motherboard. Besides, Coppermine on i840 is making Athlon’s life a whole lot harder …