As a company, Rambus hasn’t exactly endeared itself to the PC industry in the past few years. Legal patent and licensing wrangling, high RDRAM prices, and technical glitches have made Rambus infamous. Finally, after sticking with its memory partner for quite a while, Intel saw fit to begin distancing itself from Rambus late last year. The two companies still work together, but Intel has slowly and deliberately begun pursuing an alternate course by making SDRAM chipsets available for the Pentium 4. Since then, much of the market has been waiting for the Pentium 4 to transition to a new, 478-pin socketand, inevitably, away from RDRAM.
The 845 chipset, released today, is a first step. This flavor of the 845 supports good ol’ PC133 SDRAM. It’s not the fastest memory around, but it’s dirt cheap, which is important if the Pentium 4 is to replace the Pentium III in the high-volume corporate market. A version of the 845 chipset capable of using faster DDR SDRAM is in the works, but Intel says it won’t arrive until after the first of the year.
Meanwhile, feisty chipset manufacturer VIA has been selling Pentium III and Athlon chipsets with DDR support for some time now. Soon, Pentium 4 motherboards based on VIA’s Pentium 4 chipset will start hitting the shelvescomplete with support for DDR SDRAM. That fact alone might be enough to annoy Intel just a little bit, but what’s really got Intel up in arms something else: VIA has refused to pay for a license for the Pentium 4 bus, instead insisting that a cross-licensing deal between Intel and a company VIA acquired is sufficient. Last week, Intel filed suit against VIA for patent infringement.
Who says chipsets aren’t exciting?
I don’t pretend to know who’s right and who’s wrong in the VIA-Intel patent dispute, but the two companies’ SDRAM chipsets for the Pentium 4 are definitely worth a look. The 845 is Intel’s platform for pushing the Pentium 4 into the mainstream market. If the 845 serves its purpose, the transition away from the Pentium III will happen in earnest, and soon. (In fact, Intel will soon stop taking new orders for the PIII.) For VIA, the P4X266 represents an opportunity to capture a big chunk of the high-end and mainstream Pentium 4 chipset markets.
Let’s see how they compare to Intel’s RDRAM-driven 850 chipset.
The 845 chipset, code-named Brookdale, begins life supporting only PC133 SDRAM. Intel claims DDR isn’t quite ready for volume production yet, and they’ve proposed addenda to the JEDEC DDR200 and DDR266 specifications. The major concern, according to Intel, is AC overshoot/undershoot issues.
Now, whether these updates to the DDR spec are driven primarily by politics or by technical concerns, I can’t tell you. I’m sure it’s quite likely Intel’s suggestions will improve things. I also imagine Intel would prefer not to be perceived as following AMD here. You can decide that question for yourself. Honestly, I don’t know enough to make that call. I do know this: on the Athlon platform and with VIA’s P4X266, DDR SDRAM works quite well.
Anyhow, for now, the 845 chipset supports only PC133 SDRAM, which offers a peak of 1066MB/s of bandwidth. That’s quite a bit less than the 3.2GB/s peak throughput of the 850 chipset’s dual RDRAM channels. The 845 compares better when the total amount of memory counts; it supports up to 3GB of SDRAM, while the 850 supports up to 2GB of RDRAM. Other highlights of the 845’s memory controller include:
- A write cache to prevent conflicts with critical memory reads
- An IOQ depth of 12
- 24 open memory pages
- Memory refreshes are completed during low-priority memory activity to maximize bandwidth
Beyond the memory controller, however, the 845 is extremely similar to the 850. The two “north bridge” or memory controller hub (MCH) chips share a lot of the same logic: the AGP interface, front-side bus, and Intel Hub Architecture interconnect. The “south bridge” chipor, in Intel’s nomenclature, the I/O controller hub (ICH)is the same between the 845 and 850 chipsets: Intel’s 82801BA ICH2 chip.
Like the Intel 845, VIA’s P4X266 supports PC133 SDRAM. In fact, feature-wise, the PX266 seems to do pretty much everything the 845 chipset can do, plus one big addition: the P4X266 supports DDR memory right now. Up to 4GB of it, in fact. That gives the P4X266 a theoretical peak of 2.1GB/s of memory bandwidth. Combined with DDR SDRAM’s other advantageslower prices, lower latenciesthe P4X266 may give Intel’s high-end 850 chipset a run for its money.
Let’s review a few of the similarities between the P4X266 and Intel’s P4 chipsets, because the similarities are striking. Both have the requisite P4 bus, of course. Both have AGP 4X, and both have dedicated 266MB/s interconnects between north and south bridge chips. Both hang the PCI bus off of the south bridge, and they support the same I/O standards: ATA-100, USB, and AC’97. Both use Intel’s LPC interface to talk to legacy devices like keyboards, so neither chipset natively supports ISA slots. Motherboard chipsets are complex beasts nowadays, yet these two companies have achieved feature-for-feature parity.
The P4X266 is actually very similar to VIA’s KT266/KT266A chipset for the Athlon. In fact, beyond the CPU bus interface, the P4X266 and KT266A are probably almost identical. Like the 845/850, the KT266A and P4X266 share the same south bridge chip. In this case, that’s VIA’s 8233 chip.
VIA does have one more trick up its sleeve for the P4X266: due out before long is the P4M266, essentially the same product, but with an integrated S3 Savage4 graphics controller. For low-cost OEM solutions, that could make VIA’s P4 chipset quite popular.
As ever, we did our best to deliver clean benchmark numbers. Tests were run at least twice, and the results were averaged.
Our test systems were configured like so:
|AMD 760/Athlon||Intel 845||Intel 850||VIA P4X266||Intel 815EP/PIII|
|Processor||AMD Athlon 1.2GHz||Intel Pentium 4 1.6GHz||Intel Pentium 4 1.6GHz||Intel Pentium 4 1.6GHz||Intel Pentium III 1.2GHz|
|Front-side bus||133MHz (266MHz DDR)||100MHz (400MHz quad-pumped)||100MHz (400MHz quad-pumped)||100MHz (400MHz quad-pumped)||133MHz|
|Motherboard||Gigabyte GA-7DX||Intel D845WN||Intel D850GB||VIA P4X266 reference||Intel D815EEA2|
|Chipset||AMD 760/VIA hybrid||Intel 845||Intel 850||VIA P4X266||Intel 815EP|
|North bridge||AMD 761||82845 MCH||82850 MCH||P4X266 north bridge||82815 MCH|
|South bridge||VIA VT82C686B||82801BA ICH2||82801BA ICH2||VT8233 south bridge||82801BA ICH2|
|Memory size||256MB (1 DIMM)||256MB (1 DIMM)||256MB (2 RIMMs)||256MB (1 DIMM)||256MB (1 DIMM)|
|Memory type||Micron PC2100 DDR SDRAM CAS 2.5||Infineon PC133 SDRAM CAS 2||Samsung PC800 Rambus DRAM||Micron PC2100 DDR SDRAM CAS 2.5
|Infineon PC133 SDRAM CAS 2|
|Graphics||NVIDIA GeForce3 64MB (12.41 video drivers)|
|Sound||Creative SoundBlaster Live!|
|Storage||IBM 75GXP 30.5GB 7200RPM ATA/100 hard drive|
|OS||Microsoft Windows 2000 Professional|
|OS updates||Windows 2000 Service Pack 2, Direct X 8.0a|
Notice that we’ve taken the unusual step of testing the P4X266 at CAS 2 settings, though we were only using memory rated at CAS 2.5. Because this P4X266 board is a pre-production reference motherboard from VIA, we were willing to take some suggestions from VIA this time around. VIA recommended using Micron DDR memory, which is only available in CAS 2.5-rated form, and running it CAS 2 timings. We were curious to see just how fast the P4X266 could be, so we gave it a shot. Turns out only one of our three Crucial (Micron) DIMMs would run stable on the P4X266 reference board. However, since this is a pre-production sample we’re dealing with, we let it slip and tested at CAS 2 with that one DIMM. We will have a production P4X266 board from Shuttle to test very soon, and the production board won’t get such cushy treatment.
Note, also, that the Athlon system (built from production components) we included for reference was tested with Micron RAM, but only at CAS 2.5 settings. No doubt the Athlon’s memory performance would be betteralong with a slight bump for everything elseif it were tested at CAS 2 timings. And yes, in an ideal world where time and money flowed like water, we’d have re-tested the Athlon at CAS 2. In that same world, we’d have dancing girls to feed us candy and entertain us while we run benchmarks, too. We would also have a pet monkey. His name would be Earl.
If we did have a pet monkey, Earl would be explaining to you at this point that the Athlon and Pentium III systems were included here simply for the sake of reference. This is not a CPU performance comparison, and the fact the Pentium 4 is at 1.6GHz and the other two processors are at 1.2GHz doesn’t mean anything terribly important in the grand scheme of things. In truth, the slightly finicky VIA reference mobo dictated the speed at which we were able to test the P4, and we chose 1.2GHz as a reasonably comparable speed for the other two CPUs. If this clock speed disparity makes your head hurt, or if you want to see a full-blown comparison between these different processors, I suggest you read my Pentium 4 2GHz review.
The test systems’ Windows desktops were set at 1024×768 in 32-bit color at a 75Hz screen refresh rate. Vertical refresh sync (vsync) was disabled for all tests.
We used the following versions of our test applications:
- SiSoft Sandra Standard 2001.3.7.50
- Compiled binary of C Linpack port from Ace’s Hardware
- ZD Media Business Winstone 2001 1.0.1
- ZD Media Content Creation Winstone 2001 1.0.1
- LAME 3.70
- SPECviewperf 6.1.2
- 3DMark 2001 Build 200
- Quake III Arena 1.17
- Serious Sam v1.02
- ScienceMark 1.0
- Sphinx 3.3
Memory latency numbers are drawn from the Cachemem memory latency tests embedded in ScienceMark. We chose results from a 2MB matrix size, with a 512-byte stride between reads. All the tests and methods we employed are publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.
Obviously, the biggest difference between these chipsets is memory support, so let’s cut right to the chase. We’ll start with SiSoft Sandra’s memory tests, which measure bandwidth using a modified, more intensive version of Stream. These tests are a measure of memory performance onlyand the Sandra tests, of bandwidth in particularso don’t get too hung up on them before you see results from other benchmarks.
It’s no surprise that the 850 chipset with dual RDRAM channels comes out on top here, but there are a few surprises beyond that. First, the P4X266 does especially well, pushing around about 1100MB per second. That means DDR SDRAM is realizing over half of its bandwidth potential here, while the RDRAM system comes up shy of 50% of its 3.2GB/s theoretical peak. Next, the PC133-based 845 system actually delivers more bandwidth in Sandra’s integer test than the DDR-equipped Athlon system. That situation changes in the floating-point test, but obviously the combination of the Pentium 4 and 845 chipset is able to make the most of PC133’s available bandwidth.
Finally, the Pentium III system, also based on PC133 SDRAM, doesn’t deliver nearly as much throughput as the 845. Although the Pentium III “Tualatin” processor is equipped with hardware prefetch logic, it just can’t keep up with the P4 here. No doubt the PIII’s 133MHz front-side bus is a handicap.
Now let’s look at these same numbers in a little bit different way using Linpack.
Linpack graphs aren’t the easiest to read. Take a second to look at the axis labels, though, and you’ll get it. We’re measuring processing throughput here, in megaflops, for data matrices of different sizes. The small matrices, on the left half of the graph, fit into the processors’ L1 and L2 caches, so processing throughput is high. On the right half of the graph, where the matrices are too large to fit into the caches, performance drops. It’s there, at the larger data sizes, where we get a better sense of main system memory bandwidth. Overall, the shape of the graph gives us a nice visual picture of how a system’s tiered memory architecture performs. Concentrate on the Pentium 4 systems, and you’ll see identical throughput up to about 192K. Once we fully leave the confines of the P4’s 256K L2 cache, the lines for the three chipsets diverge. The RDRAM-based 850 delivers the most sustained bandwidth, but the DDR-driven P4X266 isn’t far behind. The numbers for the PC133-equipped 845 system are much lower, but they still overlap with the numbers for the Athlon DDR rig. All in all, these results are fairly similar to the Sandra results above.
Bandwidth is only one component of memory performance, however, and not always the most important one. There’s also latency, or the lag time involved in memory accesses, to consider. Here’s how things stack up in terms of latency:
Now we can see how the Pentium 4 chipsets stack up in a different sort of memory performance, and it’s a very different picture. The 850 chipset with RDRAM comes in last, well behind the P4X266, and even behind the bandwidth-starved 845. Latency is the Achilles’ heel of RDRAM, and it’s much of the reason Rambus-based systems don’t always deliver the real-world performance their theoretical bandwidth peaks seem to promise. RDRAM’s high clock speeds and narrow, 16-bit interfaces deliver lots of data, but lag times are a little higher than they are for SDRAM. Surprisingly, both of our non-P4 comparison systems turn in much lower latency numbers, though the Pentium 4 beats them soundly in bandwidth tests. In general terms, latency and bandwidth are interrelated; low latencies lead to high bandwidth and vice-versa. (Think of ping times on a 1.5Mbps connection versus a 33Kbps connection, and you’re getting the idea. RDRAM is a notable exception, of course.) In this case, however, it’s possible the Pentium 4’s higher memory latencies are related to its high memory bandwidth scores. The Pentium 4’s liberal use of hardware prefetch logic may be saturating the memory bus, driving up memory latencies.
Whatever the case, memory performance is only part of the overall performance picture. For our purposes, it’s most important to note that the P4X266 has a particularly nice combination of memory bandwidth and access latencies. The 845 is slower both in terms of bandwidth and latency.
Now the fit hits the shan. Biz Winstone tests performance in general office applications, where the 845 is likely to make its living. This test is also usually a good indicator of overall system performance, especially where “light use” patterns are concerned (in other words, not gaming, heavy computation, or multimedia).
Only three Winstone points separate the three P4 chipsets. The P4X266 shows its strength by turning in the best overall score, but the 845 also does reasonably well, considering. Content Creation Winstone 2001
Content Creation Winstone is a little more up our alley around here. It tests audio and image processing apps, where performance is at more of a premium.
This time around, the 850 comes out on top, but just barely; the P4X266 is right there with it. The 845 again takes up the rear, but here the performance gap is wider. The Pentium 4’s already relatively low clock-for-clock performance in and of itself isn’t necessarily a bad thing, but the 845 makes the situation even more pronounced. Those who buy computers solely on clock speed ratings could miscalculate badly in evaluating 845-based systems. LAME MP3 encoding
The LAME MP3 encoder is at the center of my digital world. Let’s see how our contenders stack up when converting a 50+ megabyte WAV music file into an MP3.
LAME isn’t a particularly memory-intensive test, and it shows. The difference between the three Pentium 4 chipsets here is only 1.5 seconds.
Quake III has long been the realm of the Pentium 4, and the P4’s dominance in Q3 has often been attributed to its high memory bandwidth. Let’s see if that theory holds up.
Give the Pentium 4 PC133 memory, and it starts to look pretty ordinary in Q3A, as the 845 results show. The P4X266, meanwhile, is right up there with the 850 yet again. Serious Sam
Serious Sam, another OpenGL-based first-person shooter, gives us a different look at its performance test by allowing us to plot the frame rates over time. Here’s how they compare:
As in Quake III, the P4X266 is just below the 850, while the 845 runs over 10 frames per second slower. 3DMark 2001
3DMark’s DirectX 8-based tests stress a system in some different ways. Most scenes are absolutely loaded with triangles, and advanced graphics features, including AGP performance, are on prominent display.
The P4X266 manages to win this one outright, besting the RDRAM-based 850 system. The performance penalty with the 845 and PC133 SDRAM is severeover 500 points in 3DMark.
Viewperf tests proficiency in workstation-class 3D graphics applications. It likes fast memory, and it stresses a system’s AGP implementation quite a bit, which makes it a nice test for comparing chipsets.
In most of the tests, the 850 chipset can’t be beat. However, the P4X266 isn’t far behind anywhere along the way. As we’re learning to expect, the 845 brings a substantial performance hit. I should note, at this point, that in almost all of our 3D tests, the 845/PC133 system runs very close to the Pentium III 1.2GHz. That’s important, because the 845 exists primarily to allow the Pentium 4 to step in and replace the Pentium III in Intel’s high-volume markets like corporate desktop PCs. The P4/845/PC133 trio isn’t a great workstation platform, but it’s at least as good as a Pentium III at such tasks.
The Sphinx speech recognition tests came to us via Ricky Houghton, who works in the speech recognition effort at Carnegie Mellon University. The tests are based on Sphinx 3.3, which is an advanced system that promises greater accuracy in speech recognition. The goal, which no system we’ve tested yet has reached, is processing speech with Sphinx 3.3 at a rate better than real time. You can find source for Sphinx at SourceForge.org.
Sphinx seems to rely heavily on two things: high memory bandwidth, and Intel’s SSE instructions, so it’s well suited to the Pentium 4. Here’s how Ricky explains Sphinx and its performance needs:
Sphinx is our speech recognition that CMU has been developing over the last 30 years. (Really! Many speech folks came out of here. Janet and Jim Baker, founders of Dragon for instance are from here. Several recognizers are based on the CMU system, this includes the Microsoft system, the apple system and even the Kurtzweil system now owned by L&H.) We have two systems, both open source: Sphinx 2 and Sphinx 3. Sphinx 2 is a semi-continuous HMM based system that runs in less than real time on a reasonable machine (PIII 500Mhz, 512MB ram) Sphinx 3 is running at 1.6 to 1.8 times real-time on a PIII 933Mhz machine with CAS2 133Mhz SDRAM. I ran it on a PIII 866 MHz with RDRAM and saw the system run at about 1.2 times real-time. It turns out that Sphinx 3 is memory limited, an increase in CPU speed results in very little improvements in speed, increases in memory bandwidth result in sizeable improvements in speed.
Let’s see how our chipsets affect performance in Sphinx.
The P4X266 moves us just a teensy bit closer to real-time speech recognition with Sphinx 3.3. That’s quite a coup for DDR SDRAM and the PX4266. No doubt whichever next-generation memory product arrives for the P4 first, DDR 333 or PC1066 RDRAM, will be the first to take Sphinx 3.3 into the Promised Land.
On to Tim Wilkens’ computational benchmark, ScienceMark. This suite of tests measures computational ability by running some well-known (in the right circles) scientific equations. Like 3DMark, ScienceMark then spits out a composite number denoting a system’s overall score in the suite.
Here’s how our contenders fared:
It’s a familiar story by now; the DDR and RDRAM chipsets are close to one another, and the 845 is back a ways. Now let’s break out some of the individual tests and see how those went.
The P4X266 is faster than the 850 in two of the three tests, but falls behind by eight seconds in Primordia. The 845 is slower, but steady.
So this is what the Pentium 4 looks like with SDRAM. For all the brouhaha over DDR versus Rambus, the real-world performance difference between the two technologiesat least with these chipsetsis statistically insignificant.
Nonetheless, VIA’s achievement with the P4X266 is striking. They’ve achieved performance parity with RDRAM, and with Intel’s own native platform chipset. This from a company whose memory controllers were, not long ago, among their products’ weakest features. Assuming VIA can overcome the legal hurdles, the P4X266 stands to become the chipset of choice for enthusiast motherboards, and perhaps for mid- to high-end mainstream P4 systems. What a great thing to be able to pull a stick of memory out of a nice, fast Athlon system and stick it into a Pentium 4 box! That’s something we’ve been wanting for a long time now. The Pentium 4 doesn’t seem so weird anymore.
With PC133 SDRAM, however, the Pentium 4’s greatest performance strengths are much more muted. It isn’t true that the Pentium 4 is a terrible performer when you take away a big chunk of the memory bandwidth that it has enjoyed while coupled exclusively to the 850 chipset. The 845’s performance isn’t stellar, but it’s only about 5 to 15% slower than the 850, most of the time. Still, in some of those places where the P4 has excelled, like Quake III, the 845 exacts a sizeable speed penalty. PC enthusiasts and consumers will want to avoid the 845/PC133 combo. Both the 850/RDRAM and P4X266/DDR combinations offer much better overall performance, and AMD’s Athlon with DDR memory remains the best value.
Given how well the 845 extracts bandwidth from PC133 SDRAM, we can’t wait to see it paired up with PC2100 DDR memory when the time comes. Nevertheless, the 845 is poised for sales success now. It enables the Pentium 4 to take its place on corporate desktops, where performance isn’t paramount, and competition from AMD and VIA isn’t likely to present a problem.