Click here to print this article.
Re-Printed From SLCentral
Intel Pentium 4 Shootout
What's With All Of The Memory
When I first saw the i845 chipset based motherboards last year, I was appalled. Who would have thought that Intel would want to pair up the Pentium 4 processor with SDRAM?
As it was, I am a bit uneasy at the idea of an Athlon running on anything but DDR these days, but it wasn't so disturbing because the original Athlon processors ran on SDRAM platforms. There was never anything but an SDRAM based Slot A Athlon system.
Pentium 4s typically use RAMBUS Inline Memory Modules, or RIMMs for short.
RIMMs use a high-speed narrow, 16-bit wide serial bus for transfers. PC800 RIMMs operate at 400MHz DDR, which calculates to an effective 800MHz. This allows for a theoretical peak transfer rate of 1600 MB/s (800 MHz x 2 bytes).
The system bus of a Pentium 4 is 100 MHz, and your typical PC100 SDRAM has a peak transfer rate of only 800 MB/s (100 MHz x 8 bytes). Knowing this, I had thought that the i845 SDRAM boards wouldn't last. I honestly believed that no one in his or her right mind would buy one.
As it turns out, people actually opted for a crippled Pentium 4 running with SDRAM instead of an AMD Athlon as way to build an affordable computer.
As I write this, a stick of 512MB SDRAM sells for $135; the highest it's been in a long time. But two 256 MB RIMMs (got to get two because today's RIMM based chipsets install RIMMs in pairs) go for $185 (same vendor). That's $50 that could be used for something else...like a faster processor, maybe?
PC1600 DDR, like PC800 RAMBUS, has a theoretical peak transfer rate of 1600 MB/s (200MHz X 8 bytes), hence the name. Despite the similar peak transfer rates, RIMMs still look faster on paper than DDR. The reason why RAMBUS wins out on the memory benchmarks goes back to the ancient PC analogy of cars moving down an 8-lane highway versus a 2-lane highway. Bandwidth, bandwidth, bandwidth.
Pentium 4 Meets DDR
Late last year, Intel introduced a DDR version of the i845 chipset and SiS revealed the 645, which can support SDRAM and DDR on the same board (but not used at the same time). Other than the price of the CPU itself, this levels the playing field a bit more between Intel and AMD.
So now, just as AMD processors can use SDRAM or DDR and even either on the same motherboard, the same is true for Pentium 4, only with the added bonus of having a RAMBUS solution as well for those with the extra flow to justify the slight speed increase.
How slight of a speed increase? That's what we're here to find out.
In this comparison, we put the same 1.9 GHz Pentium 4 on an Intel D845WN (SDRAM), Intel D845BG (DDR), Biostar M7SXF (DDR) and Intel D850MV (RAMBUS) and to give some sort of median, we threw an Athlon XP 1900+ on an AMD 761 based Biostar M7MIA. All configurations used 512MB of RAM, either in the form of 2 256s DIMMs or RIMMs.
About The Benchmarks
The first benchmarks I'm going to hit you with are the Sandra memory bandwidth benchmarks for the four boards. I'm doing this and getting it out of the way because I know a lot of AMD zealots will flame me for using it in the first place.
So why use it? I just wanted to get a comparison between the SDRAM, DDR and RAMBUS with something that focuses most specifically on the memory. The 3Dmark2000 scores will prove that this benchmark alone can't possibly be used to gauge the overall performance of a CPU or it's platform.
I also included Sandra scores for CPU Whetstone, Dhrystone and Multimedia benchmarks. These primarily reflect the CPU itself and do not reflect much change due to the type of RAM used.
The final three benchmarks are derived from running standard benchmarks in 3DMark2000, 3DMark2001, and Evolva.
Evolva is a great game. About a year ago it was praised as being a contender as game of the year, but somehow got overlooked because of Quake 3 Arena and Unreal Tournament. I like using Evolva for a benchmark because it's a quick load with no set up, and the crazy graphics get everyone's attention. Due to the circumstances surrounding this article, I needed a "quick and dirty" benchmark. Besides, Evolva is a lot busier and colorful than Q3A or UT. The colors of the game looks like what would happen if Elmo ate a Picasso painting with an Orange Julius chaser and then got run over by a Volkswagen Harlequin Golf. What else should we expect from Computer Artworks, the company that brought us Organic Art.
3DMark2000, as well as 3DMark2001, was used because not only does the video card get taken into consideration (obviously we're not looking at my lame Asus V7700), but also because the lower resolution scores can be used for a relatively good estimate of a CPU and it's memory subsystems' performance. Also, the game engine used in 3DMark2000 is DirectX 7 based, so it does not make use of the pixel and vertex shaders that the GeForce has, so there is again less reliance on the video card.
The game engine used in 3dmark2000 is DX7 based, it does not make use of the pixel and vertex shaders that the GeForce3 has, so there is again less reliance on the video card.
About The Benchmarker
So what are the special circumstances surrounding this review? Well, I'm sure some of you folks are thinking, "Wow! I can't believe that Jonny has five motherboards that he can run benchmarks on all at the same time. I wonder what he's going to do with those boards when he's done?"
The D845BG board was a loaner from Intel. They wanted my impression and then I had to give the board back. The D845WN and the D850MV were borrowed from my work, as well as the 1.9 GHz processor. This is rough work, I tell you! I was only given enough time to do the benchmarks I had done and none more.
Why bother with doing a comparison at all? As you may have read in my Guru's World columns, I do this for a living. People buy the parts from my work and then make their feeble attempts and building their own PCs. Most "builds" are pretty much me putting together "test rigs" on my bench on top of small cardboard boxes so I can try to duplicate exactly what the customer is doing, only correctly. I fire the test rigs up, hold the phone up to the running machine and say, "I can make it work...why can't you?"
Well, on occasion, I'll get the person that bought the D45WN motherboard, a 1.8 GHz CPU, a couple DIMMs and brag that they have this killer PC. The way I see it is that they could've spent the same amount of money on an Athlon 2000+, a DDR motherboard and DDR RAM for just as much money! Why bother with the P4 at all? You're completely crippling it with the SDRAM anyway!
Uh oh! Now you start getting into the Intel versus AMD flame wars. AMD runs too hot and has stability issues. VIA chipsets have compatibility issues. The CPUs are too fragile and you wont warranty the chip if I crack it in two. Whaaaaa!!!
If you knew what you were doing, you would have not only an AMD Athlon based system that runs 24/7 like mine, but also have it overclocked to boot. Then again, if you knew what you were doing, you wouldn't be calling me, would you?
I have nothing against Intel. They make a good product. Even as an AMD user, a zealot if you will, I can find more at fault with the AMD platform than the Intel platform. Believe it or not, I also own a P4 1.7 GHz overclocked to 1.9, but there are two "issues" I think we can all agree on. Pentium 4s are on the average slower than their AMD counterpart.
How is it that a Pentium 4 2.0 GHz runs at the same speed as an Athlon XP 1667 MHz in most benchmarks?
It used to be that price was an issue, but if you look at what AMD is doing with their PR ratings, you'll find that the AMD Athlon XP is now MORE expensive then a Pentium 4. For example, that Athlon XP 1900+ is only running at 1.6 GHz and costs $225. A Pentium 4 that runs at 1.6 GHz only costs $170!
What's an issue now is value.
The fact remains that MHz doesn't mean squat! Forget side by side comparative results between PR ratings and MHz ratings because you're bound to find a few benchmarks that the Pentium 4 will beat the Athlon XP at if you look at two CPUs of the same "rating", but if you took that $225 1.6 GHz Athlon XP (a model 1900+) and put it up against a Pentium 4 that costs the same, a 1.8 GHz, the Athlon will wipe the floor with that P4.
So it seems I've gone off on some tangent I can never recover from. The fact is this is my defense testimony for throwing two AMD benchmarks in the mix. For those of us who use an Athlon and were curious as to how the numbers would compare. Mind you, the XP 1900+ in the chart is "only" a 1.6 GHz processor. That CPU is running on a Biostar M7MIA board, which has an AMD 761 Northbridge.
All of the DDR boards (the Intel D845BG with i845 chipset, the Biostar M7SXF board with the SiS 645 chipset and the Biostar M7MIA) boards were tested using a pair of Samsung 256MB CAS2 PC2100. The Intel D845WN with the i845 chipset used a pair of Micron PC133 CAS2. The D850MV with the i850 chipset was tested using a pair of PC800 Samsung RIMMs.
Note that the Biostar board has lower benchmarks than the Intel DDR solution. The thing to remember about the Biostar is that the CPU, RAM and PCI bus speeds can be independently adjusted to do some insane overclocking mojo when PC2700 RAM is used, but that's another review altogether.
All of the boards were tested using a Maxtor 80GB ATA133 drive running Windows XP. The video card is an Asus V7700 Pure (GeForce 2 GTS 32MB DDR).
Reviewing The Benchmarks
As usual, Sandra memory benchmarks favor the RIMMs by giving it a score double that of the SDRAM. The other scores are what were to be expected, except for that I thought the Biostar M7MIA would have favored better against the two Intel DDR boards. I even took all of the same components and installed them on an Iwill XP333 (with the RAM only running at 133, of course) and obtained the same scores. So this isn't user error...this time. ;)
The Sandra CPU benchmarks caught me funny. Given that we are benchmarking the same CPU, but on three different boards, I thought that all of the Pentium 4 benchmarks would be very close. They are all close, but notice that SDRAM based board inches out on the Dhrystone benchmarks. Hmm...
All of the other benchmarks show a gradual increase in speed depending on the memory bandwidth available (and the brand of chipset when it comes down to the two DDR boards).
When it comes to a compromise of price and performance, I really feel that DDR is where it's at. Of course, RAM is going to have to become faster and have greater bandwidth eventually, much like RAMBUS. This is a fact and PC2700 DIMMs isn't going to cut it as "the solution" for the next two years.
AMD insists that the bottleneck of a computer is not the RAM, but rather the CPU. I remember this from back when they were prototyping the Athlon prior to it's release and ran a Quad Athlon in NT using only 32MB of SDRAM yet getting insane benchmarks. Obviously they're doing something right as their memory benchmark scores are always low in Sandra, but all other benchmark scores are quite high.
One thing I learned is that SDRAM for a Pentium 4 really isn't too bad. Sure the memory benchmarks stink, but I really thought that the penalty on the 3DMark and Evolva benchmarks would've been a lot harsher.
I still think that if you're thinking about an SDRAM Pentium 4 solution, you either need to save up more money or buy an Athlon. That's just my 2 cents. :)
Re-Printed From SLCentral