Original Link: https://www.anandtech.com/show/465
Intel's 820 Chipset - Performance using SDRAM
by Anand Lal Shimpi on January 31, 2000 4:27 AM EST- Posted in
- CPUs
As a chipset, the Intel’s 820 solution isn’t a bad one at all. No major compatibility problems are caused because of the chipset, Intel has already released the major INF updates and patches for the popular OSes so that they will support the new chipset, and with the driver support, reliability, and performance we’re used to seeing from Intel in their chipsets, the i820 actually makes for a very stable platform.
There is one major limitation that is the Achilles’ Heel of the i820 - its exclusive support for Direct Rambus DRAM (RDRAM). You can even stretch it further to say that RDRAM isn’t that bad of a solution if it weren’t for its overwhelming cost. But the final result of all of this rationalization is that, for most users, the i820 platform simply isn’t a viable option because of the incredible cost associated with RDRAM.
If RDRAM were a 3D accelerator, and gamers were forced to pay anywhere from $500 up to $1000 a pop to get one of these miracle sticks, then as long as the performance was the best available, you could expect RDRAM to sell. But if it wasn’t the fastest performing thing on the market, then, of course, you would expect the exact opposite.
The latter scenario is the case right now, with RDRAM offering a very small performance improvement (if any at all) for most desktop users, the up to 400% price premium RDRAM holds over even the most expensive PC133 SDRAM just isn’t worth it and it’s a fact that’s definitely holding back the sales of Intel’s “flagship” chipset.
As we’ve mentioned in our i820 Review as well as our recently published January 2000 - i820 Motherboard Roundup, a solution to this problem was provided and although isn’t openly encouraged by Intel (they do prefer that you use RDRAM), it is being actively taken advantage of by motherboard manufacturers. Through the use of what Intel calls a Memory Translator Hub, or MTH for short, the RDRAM signals being sent from the i820’s Memory Controller Hub, or MCH, can be effectively “translated” into SDRAM requests thus allowing SDRAM to be used on an i820 motherboard.
How it Works
The Memory Translator Hub (MTH) is an optional fourth component in the i820 chipset, the three required ones being the Memory Controller Hub (MCH), I/O Controller Hub (ICH) and Firmware Hub (FWH).
The 82805AA MTH is actually an option on both i820 and i840 motherboards, and its purpose is identical regardless of what chipset it is paired with. In order to keep things simple, we will only discuss the implementation of an MTH on an i820 motherboard.
The data bus present between the i820’s MCH and the CPU is a 64-bit wide bus running at your FSB speed, which is at least 100MHz on an i820 motherboard. The job of the MCH is to handle the memory transactions that occur between the RDRAM that would normally be present on an i820 board and the CPU. On one side of the MCH, we have the 64-bit data bus that we normally refer to as the Front Side bus, and on the other side of the MCH we have the 16-bit memory bus that provides for the interface to the RDRAM slots on your motherboard. Keep in mind that this 16-bit memory bus can be running at speeds of up to 400MHz while transferring on both the rising and falling edges of the clock. Let’s take a step back and analyze exactly what’s going on here before we proceed to complicate things.
Let’s say that you start out at point A and just entered an 8-lane highway where the maximum speed is 60 mph (the Front Side Bus), you then pass under bridge (the MCH) where, on the other end, the road emerges as a 2-lane highway where the maximum speed is 240 mph (the RDRAM memory bus). As you can imagine, crossing this bridge would create quite a bit of a traffic jam, trying to squeeze 8 lanes of cars into an emerging 2 lane highway, but the pace should definitely pick up on the other end of the bridge.
Now that we’ve simplified the existing i820 motherboard design, let’s go ahead and complicate it a bit by adding in the Memory Translator Hub that we started to talk about. The MTH works by intercepting the RDRAM requests over the memory bus and translates them into SDRAM requests fed to up to 2 DIMM slots located on the other end of the MTH. Let’s plug this into our highway example from earlier and see how congested things get with the introduction of the MTH.
Let’s say you emerge from the bridge on a 2 lane highway where the maximum speed is 240 mph (the RDRAM memory bus) but then you’re immediately fed into another bridge (the MTH). On the other side of this bridge you have another 8-lane highway where the speed has been decreased to 60 mph yet again and you finally end up to where you were trying to get to, point B. As you can tell, going from point A to point B in this complicated system of highways and bridges isn’t the most efficient in the world, and if you take into account going back and forth from point A to point B and from point B to point A, then this system has no efficiency.
What this is all building up to is how the Memory Translator Hub actually works; it intercepts the RDRAM memory calls and translates them into SDRAM requests and then does the exact opposite before sending the data back to the MCH. What this is in explanation is preparing you for is the end result, which is a fairly significant drop in performance because of this added translation stage.
Other Quirks
There are some other interesting facts about the 82805AA MTH that must be taken into account when considering a motherboard outfitted with one.
Only 4 RAS Lines are SupportedThe MTH itself can only address up to 4 RAS lines, what this translates into is a realistic maximum of two DIMM slots per MTH present on the motherboard. If a single MTH can address up to 4 RAS lines then why does this only result in support for 2 DIMM slots?
You’ll first have to take note of the use of the word “realistic” in that statement. The reason that 2 DIMM slots is a realistic maximum is because not all DIMMs take up a single RAS line, only single sided DIMMs do. If you have a double sided DIMM, such as a 128MB DIMM, then it occupies 2 RAS lines and thus a single MTH would only be able to address 2 of those DIMMs.
ASUS has outfitted their P3C-2000 with 4 DIMM slots yet they only make use of a single MTH. But if you read their documentation carefully, you can’t use 4 double sided DIMMs in the P3C-2000 at the same time.
We alluded to the thought of using more than one MTH on a motherboard in order to support more DIMM slots, however this is only possible on an i840 motherboard. The reason behind this is because on the i840 chipset you have two memory channels coming from the MCH, and thus can have one MTH per memory channel. With the i820 chipset, you only have a single memory channel and you can’t split that single memory channel up into two separate channels for use with two MTHs.
No ECC SupportWhile you can use ECC SDRAM on an i820 motherboard with a MTH, the MTH won’t take advantage of the ECC support provided by your SDRAM. While this probably won’t be an issue for most i820 motherboard owners, if you are planning to use an i840 motherboard in a workstation or server environment, this may be an important loss.
SDRAM ClockSince your SDRAM obviously can’t operate at the incredible speeds that RDRAM does and since it would be too expensive for a motherboard manufacturer to include a separate clock generator just for the MTH, the MTH derives the SDRAM clock from the FSB frequency based on a multiplier (sound familiar?).
The MTH keeps the SDRAM running at either a 1/1 ratio with the FSB or a 3/4 ratio with the FSB. The purpose of this is to allow for the use of the 133MHz FSB for 133MHz FSB CPUs while keeping the memory at the PC100 specification of 100MHz. The reason for this is because Intel obviously does not support/acknowledge the VIA developed PC133 specification as a real memory specification and thus designed the MTH with PC100 SDRAM in mind.
PC133 SDRAM will work on an i820 motherboard, you just won’t be running the SDRAM at its 133MHz theoretical maximum.
The ASUS P3C-2000 allows for the manual adjustment of the MTH clock divider, however if an FSB frequency of 133MHz or above is chosen, the divider defaults to 3/4 and won’t let you adjust it.
CRIMMs
If your motherboard features both RIMM and DIMM slots, you naturally can't occupy both, but if you're using the DIMM slots then all of your empty RIMM slots must be filled with Continuity RIMM modules (CRIMMs). Motherboard manufacturers that feature both RIMM and DIMM slots generally ship their boards with enough CRIMMs to fill all of the RIMM slots.
Heat
The last point of concern when implementing a MTH on a motherboard is that during the translation process the MTH naturally generates quite a bit of heat. This has forced most motherboard manufacturers to place a heatsink on the MTH in their designs.
The Motherboards
We had a number of options to choose from when selecting motherboards for this comparison, however we ended up choosing different motherboards for our i820 + RDRAM and i820 + SDRAM test beds simply because of flexibility reasons.
As we proved in our i820 Motherboard Roundup, the performance difference between all i820 motherboards is negligible and thus wasn’t an issue for us picking different motherboards for the two main test beds.
For the RDRAM equipped i820 test bed we chose the AOpen AX6C-L because of its ability to manually adjust the RDRAM clock multiplier. This allowed us to keep the RDRAM transferring at its peak rate of 800MB/s even at the overclocked 150MHz FSB setting we used.
For the SDRAM equipped i820 test bed we chose the ASUS P3C-2000 because of its ability to manually adjust the SDRAM clock multiplier. This allowed us to keep the SDRAM frequency as close to 133MHz as possible regardless of the FSB setting, however because of the fact that the motherboard would lock in the value at 3/4 if the FSB was above 133MHz, we weren’t able to fully level the playing field.
For the sake of comparison we included our Overclocking the FC-PGA BX and VIA Apollo Pro 133A test beds as well.
The Test
Windows 98 SE Test System |
||||
Hardware |
||||
CPU(s) |
Intel
FC-PGA Pentium III 550E |
|||
Motherboard(s) |
AOpen
AX6C-L
|
ABIT
BE6
|
Tyan Trinity 400 |
ASUS
P3C-2000
|
Memory |
128MB Samsung PC800 RDRAM provided by Mushkin |
128MB
PC133 Corsair SDRAM
|
||
Hard Drive |
IBM Deskstar DPTA-372050 20.5GB 7200 RPM Ultra ATA 66 |
|||
CDROM |
Phillips 48X |
|||
Video Card(s) |
NVIDIA RIVA TNT2 Ultra 32MB (default clock - 150/183) |
|||
Ethernet |
Linksys LNE100TX 100Mbit PCI Ethernet Adapter |
|||
Software |
||||
Operating System |
Windows 98 SE |
|||
Video Drivers |
|
|||
Benchmarking Applications |
||||
Gaming |
GT
Interactive Unreal Tournament 4.04 UTbench.dem |
|||
Productivity |
BAPCo SYSMark 2000
Ziff Davis Content Creation Winstone 2000 |
All gaming
tests run at 640 x 480 were run in 16-bit color to stress the CPU
All gaming tests run at 1024 x 768 were run in 32-bit color to represent real
world play
Content Creation Winstone 2000 provides us with a good real world simulation of a multitasking environment. As you can see by the chart, the i820 equipped with SDRAM is about 3 - 6% slower than the i820 using RDRAM and about 8% slower than a regular BX with SDRAM.
The performance of the i820 with SDRAM is about on par with that of the VIA Apollo Pro 133A using SDRAM, while in some cases the 133A manages to pull ahead courtesy of its 133MHz memory bus frequency whereas even with the 150MHz FSB setting, the i820 w/ SDRAM ran the memory bus at 112MHz.
SYSMark 2000 paints a slightly different picture than CC Winstone 2000 primarily because of the nature of the benchmark. SYSMark 2000 focuses on running a single application at a time whereas CC Winstone 2000 is more of a multitasking benchmark.
Under SYSMark 2000, the i820 + SDRAM combo is almost 10% slower than the i820 + RDRAM combo. The performance difference between the BX and i820 + RDRAM platforms is negligible thus helping to reaffirm the statement that for the average desktop user, the incredible cost of RDRAM just isn't worth it.
On a clock for clock basis the Apollo Pro 133A (694X) test bed seemed to pull slightly ahead of the i820 + SDRAM setup. Considering the cost of an Apollo Pro 133A motherboard is considerably less than the ASUS P3C-2000 used here for the i820 + SDRAM setup, this isn't too promising of an argument for the i820 chipset's SDRAM support.
The MTH heavily penalizes the i820 + SDRAM setup here as it is beaten by every since competitor, BX, 133A and i820 + RDRAM, on a clock for clock basis. The performance hit the i820 takes from the MTH is so bad that even at 825MHz the FC-PGA on a BX running at 682MHz has no problem beating it.
UnrealTournament definitely appreciates the added memory bandwidth provided for by RDRAM thus giving the top 4 spots away to the 4 fastest RDRAM setups.
Once again the Apollo Pro 133A seems to have no problem beating the i820 + SDRAM on a clock for clock basis although this time it takes a 733MHz clock speed on the 133A to beat out the 825MHz setup on the i820 + SDRAM combo.
We have a similar situation here at 1024 x 768, but the standings are much closer.
Conclusion
The idea of putting SDRAM on an i820 motherboard is definitely a good one and a very practical one, at least in theory. But in practice, through the use of Intel's Memory Translator Hub, it is far from the most efficient idea. Unfortunately there is really no way around this other than to go after a motherboard that uses either the BX, the Apollo Pro 133 or 133A chipset.
As a complement to an overclocked FC-PGA CPU, an i820 motherboard with SDRAM does become an interesting option but when you consider that a 500E running at 750MHz (150MHz FSB) on an i820 with SDRAM is slower than the same CPU running at a lower clock speed of 620MHz (124MHz FSB) on a BX board it really makes you want to reconsider throwing out your trusty BX motherboard.
We've said it before and we'll say it again, the BX platform is still the most attractive platform in terms of stability and performance for the newer Pentium IIIs based on the Coppermine core. Even with SDRAM the i820 loses its appeal because of the heavy performance penalties associated with using a MTH in order to gain SDRAM support.
If you must have official 133MHz FSB support, the VIA Apollo Pro 133A is a much better and much more affordable chipset option than the i820 + SDRAM. Unfortunately we are still waiting for the kinks to be worked out of most of the current Apollo Pro 133A based motherboards as well as for others to finally hit the streets.
If there only were a better 133MHz FSB chipset on the horizon... ;)