Intel i820 Chipset Review — THG.RU

Introduction

The long anticipated ‘Camino’ or i820 Chipset still isn’t released! Intel’s first target was for June 99 and then was pushed forward to September 99. Just days before Intel’s planned launch, September 27, 1999, an issue with memory data integrity when more than 2 banks of RDRAM are installed puts a quick stop to the products launch. Tom did the first preview of the i820 chipset set back in June, Preview of Intel’s Upcoming ‘Camino’-Chipset. Intel was so kind to provide me with a newer sample of the chipset on their VC820 i820 based motherboard. To make my testing more interesting, I’ve included Intel’s tried-and-true 440BX and VIA’s Apollo Pro 133 chipsets into my comparison. Given the new features provided by the i820 chipset, I did my best to try and extract the performance benefits each feature has to offer.

What does the i820 bring to the Table?

The i820 includes several new features that were implemented to increase bandwidth across the platform. Here is my list of the most important features implemented into the i820 chipset.

HUB Architecture

The i820 is based on the same HUB architecture as the already shipping, low-end i810 chipset. Basically these two chipsets architecture’s are identical except the i820 has a 4X AGP bus (without video integration), 133MHz FSB capability, and RDRAM support. We did two comprehensive reports on the i810 chipset in April 99: Whitney, Intel’s 810 Chipset — Part I and Whitney, Intel’s 810 Chipset — Part II. To assist in clarifying the new HUB architecture of the i820, I’ve included a diagram for it and the North / South Bridge based i440BX chipset. As they say: «Pictures are worth a thousand words.»

These diagrams demonstrate the i820’s advantage with its new architecture. Due to the MCH having direct access to the memory bus, the slow PCI bus does not limit the i820. The i820’s UDMA66 support provides twice the bandwidth as the 440BX’s UDMA33. However, even the fastest ATA-66 drives can only sustain ~13MB/s transfer rates.

‘MCH’ and its Buses

The 82820 chip is named MCH, which stands for ‘Memory Controller HUB’. The MCH includes a couple of important buses.

The Memory Bus — 133MHz, 128 Bit = 1.6GB/s

The memory bus in combination with RDRAM is twice that of the BX chipset. However, let’s not forget that the CPU can only take advantage of 33% more memory bandwidth, and this is only the case if the CPU runs at 133 MHz FSB. Also, because AGP has moved to 4X it requires increasing the memory bandwidth so it can keep up with the AGP bus. The new RDRAM or Rambus memory does have a much higher bandwidth, but it suffers from higher latencies vs. PC-100 and PC-133 SDRAMs.

The ‘Interlink» between the MCH & ICH — 133 MHz, 2x, 8 bit = 266 MB/s

The MCH and ICH are able to directly communicate at 266MB/s. This is much faster than the 440BX’s (133MB/s) communication between the North and South Bridge via the PCI-bus.

The AGP 4X bus — 133MHz, 128 Bit = 1064MB/s

The new AGP 4x bus provides much more bandwidth for the AGP graphics adaptor, twice as much bandwidth as the AGP 2X on the 440BX.

The ‘ICH’

Right underneath the ‘MCH’ you can find 82801AA, the ‘ICH’ = ‘I/O Controller Hub‘. This part of the chipset talks to all the PCI-devices over the PCI bus, to the EIDE hard drives over the ATA66 interface, to external devices over the USB, to a low priced modem/audio-codec via AMR and finally to the ‘FWH’ = ‘Firmware Hub’. The ICH is the place where all the outside communications run together, forwarded to main memory, the CPU or the AGP graphics controller over this new 266 MB/s bus. You could say that the ‘ICH’ replaces what used to be called ‘South Bridge’.

The FWH — 82802

Behind the name ‘FWH’ = ‘Firmware Hub‘ you’ll find a chip that’s not much else than a 4 Mbit EEPROM plus a tiny bit of active silicon. The EEPROM contains the motherboard BIOS and the active silicon is a random number generator. This random number generator, otherwise known as the ‘RNG’, can supposedly be used by software to provide extra security during computer related transactions (e.g. ECommerce).

RDRAMs — Rambus Memory

Say goodbye to the old familiar SDRAM modules. The i820 chipset only supports RDRAM memory. This new memory technology provides twice the bandwidth of conventional SDRAMs. The only two drawbacks are the memory cost and slower latency. The RDRAM cost is nearly 5 times the cost of SDRAM PC-100 memory, but as SDRAM-prices rise, the difference may shrink. It still adds a whole lot of additional cost to an i820 based system though. Also, you will find that i820 based systems will always have every memory bank filled. If the memory slot doesn’t have a RDRAM RIMM module installed it must have a continuity module called a CRIMM in the memory socket. The CRIMM(s) is required to avoid any reflections with the high speed Rambus-interface.

Here is a photo showing a RDRAM RIMM module (blue) with two CRIMM or continuity modules.

Future i82x chipsets will provide support for PC-133 SDRAM but all of the memory accesses will be forced to go through the MTH, otherwise called ‘Memory Translator HUB’. I don’t have all the details regarding this new HUB but I’m expecting some performance hits due to the translation process. The i820 chipset supports various RDRAM speeds. However, not all of these options are available to every FSB frequency. Here is a table that depicts supported frequencies.

RIMM Sizes 64MB, 96MB, 128MB, 256MB

Is there a new CPU that supports 133MHz Bus?

There is a different processor that supports the i820’s 133MHz FSB. Basically, it the same old Pentium III processor that has been qualified to run with the new higher bus speed. The main difference between the current Pentium III and the new one is the bus multiplier each CPU is clocked at. Today’s Pentium III 600 runs with a FSB frequency of 100MHz and a multiplier of 6x. A Pentium III that supports the higher 133MHz FSB has a lower multiplier for a 600MHz core. With a 133MHz FSB the multiplier would be 4.5x. The new 133MHz capable Pentium III’s will be labeled 533B (for 533MHz) and 600B (for 600MHz). Other than the ‘B’ at the end of the CPU’s frequency, nothing has changed about the new, or is it old, Pentium III processors. Even though most of the current shipping processors could probably handle the higher 133MHz FSB, the problem would be the higher locked multiplier. It would be out of the question to expect a 600/100MHz part with a locked multiplier of 6x to run with a 133MHz FSB at 798MHz, but it may be a nice chance to overclock e.g. a Pentium III 450 to 600 MHz.

Platforms — Intel VC820 Motherboard (i820 chipset)

Platforms — ASUS P3B-F Motherboard (i440BX chipset)

Platforms — EPOX EP-6VBA Motherboard (VIA Apollo Pro 133)

Environment Settings

Business Application Performance — L2 cache ENABLED

To give an idea of business application performance I chose BAPCo’s SYSmark98 & Intel’s Business Application Launcher v1.2. We’ve found that SYSmark98 provides a much more reproducible score than Winstone. To provide a good comparison I put the i820 board (Intel’s VC820) against ASUS’s P3B-F (440BX chipset) and Epox’s EP-6VBA (VIA Apollo Pro 133 chipset). SYSmark98 was ran under both Windows 98SE and NT 4.0 SP5.

Looking at the SYSmark98 results under Win98 the i820 performs just about as fast as the i440BX platform. The VIA based system lags behind both the i820 and i440BX. However, Netshow Encoder and Photoshop seem to enjoy the higher memory bandwidth the i820 & VIA platform offer.

Under WinNT the SYSmark98 results show pretty much the same as the Win98 results. No major performance gain over the i440BX.

Business Application Performance — L2 cache DISABLED

I also ran SYSmark98 under Win98 with the CPU’s L2 cache disabled on each platform to make the systems performance more dependent on the main memory. Of course this introduces a performance hit, but it will also help us determine the benefit of the i820’s increased memory bandwidth.

By disabling the L2 cache the i820 really shines. This shows that the i820 and its new RDRAM (Rambus) memory do indeed have a positive effect on the business application results. However, I was completely surprised that the VIA platform was only able to outperform the i440BX in one category, even with it’s faster 133MHz FSB.

Business Application Performance — SCSI Hard Disk

By disabling the L2 cache I was able to emphasize the performance effects of the main memory. But I also thought it would be interesting to remove the onboard IDE controller from the picture also. I installed an Adaptec 2940UW SCSI adaptor and an IBM DVGS 9GB 10,000 RPM SCSI drive into each platform. I wanted to put each platform at an equal level in the disk subsystem category. Doing this would rule out and mischievous IDE bus-mastering driver benchmarking tweaks. I figured this would also point out if any one of the platforms were lagging due to their IDE interface implementation.

Using SCSI really didn’t change the performance deltas between any of the platforms. Even the i820’s integrated UDMA66 didn’t benefit the platform when using IDE.

Game Application Performance — AGP

We’ve already speculated that the i820’s AGP 4x wasn’t going to yield a huge performance gain in today’s games. I’ve included results using Expendable Demo, Quake 3 Arena Test and Shogo. Each of these games was run at a resolution of 640x480x16. I used the lower resolution to keep the video board from being the bottleneck.

In Expendable the i820 has a minuscule lead over the i440BX platform. The VIA platform is over 8FPS slower the i440BX. This proves once more that VIA’s AGP-implementation is still not up to speed with Intel’s.

Under Quake 3 the i820 and i440BX run a tight race. The VIA platform stills holds a steady last place.

Nothing has really changed even with the Shogo results. Both the i440BX and i820 platforms perform almost identically. The VIA based system still tries to play catch up. Could the VIA board’s performance be limited by their AGP implementation?

AGP 4X Performance

Since none of the game benchmarks in our test suite showed any significant performance advantage with AGP 4X I decided to throw in an AGP test that purposely shoves lots of textures through the AGP bus. The results are given using Intel’s Platform Bandwidth Test. This test is comprised of a complex 3D scene using the integration of video textures and movie playback along with some typical 3D graphics. There are two video streams (in .avi format) that are decoded on the fly and used as textures. I’ve included a screen shot of the test. In this scene the room is rotated 360 degrees while the AVI’s are playing.

In this large textured-based benchmark you can see that AGP 4X can in some circumstance offer considerable performance gains.

Does my AGP Graphic Adaptor support 4X AGP?

A lot of video boards have been claiming AGP 4X support for quite some time now. However, there really hasn’t been any platform to take advantage, let alone test this 4X AGP functionality. Anyone who already has a video board and is planning on making the expensive upgrade to an i820 based platform (when available), or is planning to purchase a board might want to make sure the product they choose supports AGP 4x. Even though a particular chipset, such as NVIDIA’s TNT2, supports 4X AGP. If 4X AGP support wasn’t implemented in the video board design, that video board may not support it. Even though 4X AGP really doesn’t show us any huge performance gains right now. If you are planning to purchase a product in the near future you might want to make sure it does. Obviously the best way to determine if your video solution supports 4X AGP is to check with the manufacturer of the board. Otherwise usually you can tell if the video board supports 4X by looking closely at the edge connector of your video adaptor. Check out this picture.

Example: Video board that supports AGP 4x

A close look at this Diamond Viper V770 Ultra shows the extra notch required for AGP 4X support. The Diamond V770 also includes a jumper that has to be set to enable 4X AGP.

Example: Video board that doesn’t support AGP 4x

This Creative Labs TNT2 Ultra board doesn’t have the required notch to support 4X AGP.

Again, this is a pretty sure shot way to determine if your board supports AGP 4x or not. But it is always best to check with the board manufacturer to get their boards exact specifications.

Game Application Performance — PCI

Looking at the results using an AGP board really didn’t show us much improvement in the i820’s new 4x AGP bus. I decided to plug in a PCI based TNT2 adapter, ASUS’s V3800 PCI. I was curious to see how the VIA platform would compare by removing AGP performance completely.

Now all of the PCI based gaming benchmarks really caught me off guard. It looks as though the i820 platform has a pretty descent advantage when the video is utilizing the PCI bus. With an ASUS V3800 PCI TNT2 Ultra based board installed in each of the platforms the i820 had a 3FPS advantage over the i440BX! When an AGP based video adapter was installed the i820 and i440BX almost scored identically in this same test.

Even under Quake 3 Arena the i820 had a superior score over the other platforms when a PCI based video adaptor was installed.

It clear that the PCI bus of the i820 has superior performance over the other platforms, at least with a PCI video adaptor. I thought that the VIA platform would come into the running with the PCI video board. I figured that by removing the AGP bus from the picture the VIA might get a little closer to the i820 & i440’s performance. It looks as though the VIA’s PCI and AGP buses are just a bit slower across the board in comparison to the i820 and i400BX chipsets.

Conclusion

Since our last dealings with the ‘Camino’ or i820 chipset it is clear that Intel was able to clean up some of their performance. In all of our tests the i820 was able to keep pace with the i440BX based platform, sometimes even outperforming it. A lot of new features and increased bandwidth have been squeezed into this newer chipset. Unfortunately, most of today’s software and hardware don’t take advantage of it and I dare to ask for the benefit of RDRAM. I guess the answer will be the same marketing-hype we’re used listening to for years: «You may not see any advantage yet, but future software will make a difference.»

I wonder how long we’ll have to wait for that, possibly longer than the short life span of Intel’s CPUs and chipsets. So far I’m still waiting to see a decent amount of software that takes advantage of the oh-so-wonderful ‘Streaming SIMD Extensions’ or even MMX. It could be that as time passes, some of the increased bandwidth will start to show through in the new HUB architecture. As new USB devices become popular, such as USB networking, some of the bandwidth advantages will become apparent. It’s about time that the PCI bus got relieved from the I/O-traffic between the chipset’s north and south brigde. I also like AGP4x, which I consider the most if not the only meaningful advance of i820.

Camino or ‘i820’ will make its way, once the latest problems got resolved. It certainly isn’t any great hardware the world has been waiting for. Intel tries to push us into a new direction called ‘RDRAM’ and the majority of the computer using sheep will obey, forcing the critical rest to follow. Do we need i820? Well, we need it just as urgently as SSE, breast implants and hair transplantations. As long as marketing wins over intelligence, people will follow Intel’s lead, will pay a premium for a Pentium III 600 over a Pentium III 550 although it only offers an unnoticeable 3% performance gain and they’ll pay lots of money for i820 and RDRAM. The show must go on, and the computer industry has gone Hollywood long ago.