We always like to look back at things we have said in the past and use that as a starting point to introducing new technology.  Keeping in line with what we’ve done historically, here is an excerpt from the conclusion to our first review of the Intel Pentium 4 from November 20, 2000:

“While it's a good idea for Intel to attempt to take the price of RDRAM out of the picture by bundling two sticks with each boxed CPU, this isn't a true solution to the problem. The solution that needs to be implemented is that the Pentium 4 needs a DDR SDRAM platform, preferably one from Intel (VIA hasn't always had the best memory performance) and it needs one before it's too late. According to Intel, the Brookdale chipset (DDR SDRAM for the Pentium 4) won't be out until the first quarter of 2002, by that time even Dell will be begging for AMD chips if there is no DDR chipset for the Pentium 4. If Intel doesn't come through with one it seems like it will be up to VIA, luckily their DDR memory controller is already sampling in Apollo Pro 266 chipsets so it’s mainly a matter of licensing the bus from Intel and implementing it in a North Bridge design.”

There are a couple of points made in the above paragraph that require some digestion.  First of all, the price of RDRAM has come down considerably since we originally published that statement.  It still has yet to achieve a price parity with SDRAM or DDR SDRAM, but it’s much cheaper than it used to be.  In spite of this, it costs approximately twice as much to purchase RDRAM as it does to purchase SDRAM/DDR SDRAM; this is especially true because of Crucial’s very aggressive pricing of DDR SDRAM.  One competing manufacturer informed us that they have to produce four times as much memory now to maintain the same profits from before the aggressive Crucial price drops.

Secondly, VIA has continued their improvement process over this past year.  Although the KT266 chipset was plagued with problems initially, today it is a viable solution for many users.  The KT266’s DDR memory controller is working perfectly fine, and although it doesn’t offer the highest level of performance, it is still respectable. 

Third, Intel’s i845-D (Brookdale with DDR SDRAM support) is still on track to be introduced in the first quarter of 2002.  In spite of sluggish Pentium 4 sales Intel has not pulled the 845-D in any closer.  Even if they did, with only DDR200 (PC1600) SDRAM support the i845-D wouldn’t offer that much of an advantage over the regular i845 with PC133 SDRAM.  Intel has yet to qualify DDR266 (PC2100) for use with their platforms. 

Fourth, in order for VIA to enter the Pentium 4 chipset market they will have to have a license for the Pentium 4 bus.  Intel was willing to grant VIA a license for the bus, but at very high costs.  Our sources gave us estimates of between $10 and $20 per chipset in licensing fees.  VIA’s chipsets rarely sell for more than $40, with $30 being the sweet spot.  With Intel taking 25% - 50% of that for licensing fees, it’s definitely cost prohibitive for VIA to enter that market.  Luckily VIA has another alternative, because of their acquisition of S3 Graphics VIA claims to have inherited licensing rights to Intel’s buses which were originally given to S3. 

The culmination of those four points has surfaced today in the form of VIA’s first Pentium 4 chipset, the P4X266.  What’s so special about this chipset and what’s so irritating to a handful of people at Intel is that the P4X266 does not support Rambus DRAM like Intel’s i850; rather, the chipset is designed for use with DDR SDRAM.

Today it’s time to find out whether RDRAM’s incredible bandwidth is necessary for the Pentium 4 to perform well or if DDR SDRAM is more than enough. 

The P4X266

It wasn't that difficult for VIA to come up with the design for the P4X266.  The North Bridge borrows the AGP 4X controller and V-Link interface from the KT266.  The DDR memory controller is taken from the updated (and presently unreleased) revision of the KT266 chipset as well.  The only unique part about this chipset is the Pentium 4 bus interface which we can't imagine took VIA all too long to implement in a design. 

As with all of VIA's present day offerings, the P4X266 is a two chip solution.  The P4X266 North Bridge is the heart and soul of this chipset.  Although it supports regular SDRAM, almost all P4X266 based solutions will take advantage of the North Bridge's DDR200/266 memory controller.  Like most VIA chipsets, the P4X266 allows for asynchronous operation of the FSB and memory bus.  This is actually very important with the P4X266 because the Pentium 4’s FSB is clocked at 100MHz (quad-pumped) while DDR266 SDRAM runs off a 133MHz clock.

Although not officially supported, the chipset will work fine with the 133MHz FSB enabled.  This is actually the same case with Intel’s 850 chipset but it’s a rarely used overclock because of the fact that it’s not that easy to get Pentium 4 processors to run at a 33% higher frequency. 

The VT8233 South Bridge is the same V-Link enabled South Bridge that is present on all KT266 and Apollo Pro 266 motherboards.  With three USB controllers the South Bridge supports a total of 6 USB ports.  Support for Ultra ATA 100 devices is present as well.  The South Bridge does feature VIA’s integrated network controller, and for the IT markets that require a brand name Ethernet provider the VT8233C South Bridge has an integrated 3Com controller.  A future version of the South Bridge will offer ATA/133 support.

The two chips are connected by VIA’s 8-bit V-Link bus that is capable of transferring at up to 266MB/s. 

There is nothing fundamentally new about the P4X266 which is why it’s not surprising that VIA was able to go from beta to production silicon so quickly.  If you’ll remember from our Computex coverage, the first time VIA had ever publicly shown and benchmarked the P4X266 was in front of right us at their Computex press conference.  Almost two months later, VIA is ready to begin shipments of the part to motherboard manufacturers.  There are even a few bigger name manufacturers that already have boards ready and should be able to offer them in a reasonable timeframe. 

VIA will be introducing a pin compatible version of the P4X266 North Bridge with integrated S3 Savage video.  This will be called the P4M266. The P4X266 supports both Socket-423 and Socket-478 processors as well as the upcoming Northwood core (0.13-micron Pentium 4).

Let’s talk bandwidth

The name of the game with the Pentium 4 has always been bandwidth and until now we’ve never been able to challenge that assumption.  The Pentium 4 was introduced with Intel’s own 850 chipset platform which only supported a dual channel RDRAM memory interface.  If Intel had it their way, all of the Pentium 4 solutions offered would be RDRAM based from the entry level platforms to the highest end.  The reality of the market has prevented Intel from doing this and has forced them to consider SDRAM as a solution for the lower price points. 

There is good reason for Intel to want to stick with the dual channel RDRAM solution that the i850 chipset offers.  With PC800 RDRAM, the two 16-bit RDRAM channels offer a combined peak bandwidth of 3.2GB/s.  This is exactly the same amount of bandwidth that the Pentium 4’s 64-bit quad-pumped 100MHz FSB offers to the processor.  In theory, an equal match of FSB and memory bandwidth would eliminate any bandwidth bottlenecks that could arise thus allowing the CPU to run at its full potential.  Another part of this theory of how bandwidth effects CPU performance is that as CPU speed increases, its dependency on a higher bandwidth memory bus will increase since it will be processing/requesting more data at any given time. 

With that said, how much is enough?  Obviously any more than 3.2GB/s isn’t going to give your CPU much of a performance increase.  The V-Link connection to the South Bridge and the AGP 4X controller will eat up a bit of the 3.2GB/s of memory bandwidth but exactly how much is very difficult to determine.  From what we’ve seen, the amount that they consume at this point is very limited. 

On the SDRAM side of things, PC133 SDRAM is obviously not enough as it can only deliver 1.06GB/s of memory bandwidth.  Even for today’s Athlon platforms PC133 SDRAM can be limiting although not nearly as much as with the Pentium 4. 

Otherwise known as PC1600 DDR SDRAM, DDR200 SDRAM offers a bit more bandwidth at 1.6GB/s but that’s still only half of what a dual channel PC800 platform offers.

The obvious and most competitive solution to the dual channel RDRAM of Intel’s 850 chipset is DDR266/PC2100 DDR SDRAM.  A single DDR266 channel offers 2.1GB/s of bandwidth.  This is currently as close as you’re going to get to the 3.2GB/s that RDRAM has been able to give the Pentium 4.

The Board and its competitors

VIA provided us with a reference board based on the P4X266 chipset for use in this review.  The reference board was very reliable during our tests and crashed only once during the days of non-stop testing.  Unfortunately the reference board would not work with our unlocked engineering sample Pentium 4 processors so we were forced to use the only locked CPU we had in the lab: a Pentium 4 1.5GHz processor.  This won't affect you but we felt it was worth explaining why we tested with a Pentium 4 1.5GHz CPU instead of the current P4 speed champ, the Pentium 4 1.8GHz.

Click to Enlarge

We compared the board to MSI’s 850Pro2 which was one of the first 4-layer i850 motherboards to hit the streets.

Finally, we managed to get our hands on an ABIT BL7-RAD through a distributor contact of ours.  The BL7-RAID uses Intel’s i845 chipset (aka Brookdale) which is a Pentium 4 chipset with PC133 SDRAM support. 

Click to Enlarge

In spite of the fact that Intel’s official i845 launch is a month away, you will be able to purchase boards well before then.  There are some floating in the channel already and you can expect that number to increase in these next couple of weeks.

The Test

Memory Bandwidth - Linpack

Our analysis starts off with a slew of memory performance benchmarks.  This is mainly to get a handle on what to expect from these three Pentium 4 platforms and exactly what's going on. 

Linpack hasn't been seen in a review at AnandTech in a little while, mainly because we had already explained what was going on in the benchmark countless times.  But now with new platforms, we were eager to dust it off and give it a try.

If you don't remember how to look at this graph from our other reviews just keep in mind that the Pentium 4 CPU only has a 256KB L2 cache.  As the data size gets larger than 256KB it can no longer fit in the CPU's on-die data cache and is forced to go to main memory.  The performance of the application drops along with this drop in memory bandwidth and increase in memory latency. 

The lines to the right of the 330KB marker are what you really want to look at.  As expected, the i850 with it's dual channel RDRAM offers by far the most bandwidth out of the three.  Interestingly enough however, the Pentium 4 makes much better use of DDR SDRAM than the Athlon platforms did in this benchmark (see our Socket-A Chipset Comparison for more details).  At the larger data sizes the P4X266 chipset offers approximately 66% of the performance of the i850, again governed by the chipset's memory bandwidth.

Conveniently enough, the i845 offers 45% of the performance of the i850 in this test.  Does this mean that the P4X266 will perform at 66% of the 850 and the 845 will perform at 45% of its older brother?  In an extremely memory bandwidth hungry application, maybe; but remember that we're still stuck in the theoretical world for now. 

Memory Bandwidth - Sandra 2001

In theory the dual channel RDRAM solution on the i850 provides about 50% more bandwidth than the P4X266's DDR memory controller when operating at full speed.  SiSoft Sandra's STREAM test actually works very well when attempting to measure bandwidth that's available to applications in real world tests, and these tests agree pretty much with what the specs state: 40% more bandwidth for the i850.  The real question is, do we need 1.6GB/s or will 1.1GB/s do just fine?

The picture doesn't change much here.

Memory Latency: Things get Interesting

While we ran Cachemem's bandwidth tests, they didn't provide us with any new information outside of what Sandra already told us.  What was quite interesting was the latency of these three memory subsystems.

There was virtually no change in latency going from the i850 to the i845.  This could mean that the 850's dual channel RDRAM solution coupled with the Pentium 4's high clock speed truly hides RDRAM's latency.  It could also mean that the i845 was purposefully crippled by Intel since in theory, regular SDRAM should have a lower latency than RDRAM. 

The reason we have to believe that the i845's SDRAM controller was purposefully crippled is evident when you look at the winner of this benchmark: VIA's P4X266.  The P4X266 achieves a 15% lower latency with DDR SDRAM than Intel's 845 can pull off with regular SDRAM.  This is very impressive indeed, especially when you take into account that historically Intel has been the one with superior memory controllers.  This same very low latency memory controller should be present in the upcoming revision of the KT266 North Bridge. 

Do we even want Intel's DDR 845 platform to come out if it will be inherently crippled?

Now that you have an understanding for how these platforms compare in memory bandwidth and latency, let's get into the applications themselves.

Business & Content Creation Performance

As usual, we'll start off with an analysis of performance under Winstone 2001. 

It doesn't take a rocket scientist to know that you don't need 3.2GB/s or even 1.6GB/s of memory bandwidth to run Word and Excel.  Remember that the fastest your hard drive can transfer data rarely exceeds 40MB/s for most IDE hard drives, and Business Winstone 2001 (like most of our daily computer usage) is heavily disk limited. 

The i850 and the P4X266 offer virtually identical performance here, which is a huge win for the Pentium 4.  It means that the cheaper P4X266 platform can be used in businesses without worry of a loss in performance. 

In spite of the heavily disk bound tests, the i845 is still about 13% slower than both of the higher bandwidth platforms; definitely not worth it.

Things don't change much under Content Creation Winstone 2001 either.  The i850 and P4X266 offer similar performance while the i845 falls behind by around 10%. 

Office Application & Content Creation Performance

SYSMark 2001 has always been much more of a memory bandwidth intensive test than Winstone 2001 which is part of the reason why the Pentium 4 does so well under it. 

SYSMark's Internet Contention Creation test is very taxing on a memory subsystem.  A large part of the test score is computed from a run of Microsoft's Windows Media Encoder while performing other tasks alongside it.  In spite of this, the 850 holds only a 2% performance advantage over the P4X266.  At the same time, the i845 falls behind by only 8%.  Indicating that this benchmark isn't as memory bandwidth intensive as we once thought; it's the Pentium 4 itself that paves the way for the high performance, not its memory bus.

The Office Productivity suite in SYSMark 2001 shows a larger gap between the 850 and the P4X266, however performance gap is still below 5%.  The 850 only has a 7% advantage over the i845 too.

What this translates into overall system performance is that the i850 is about 4% faster than the P4X266 which itself is about 4% faster than the i845.  We have yet to see the 850's 40 - 50% memory bandwidth advantage show its face.

One of Intel's most favorite tests to show off what the Pentium 4 and the i850 could do was SYSMark's Windows Media Encoder benchmark.  This test was thought to be biased in favor of Intel because of its memory bandwidth dependency.  In order to put that theory to test, we dusted off yet another benchmark and put it to good use.

Although the i850 does hold a 6% performance advantage, it's not nearly as high as we expected it to be.  It looks like that Windows Media Encoder isn't necessarily a bandwidth intensive benchmark, but took great advantage of the Pentium 4's low latency serial optimized architecture. 

IT/Enterprise Computing Performance

We imagined that increasing the amount of concurrent memory accesses would stress the theoretical bandwidth differences between the i850 and the P4X266 more than we had before so we turned to Office Bench 2001.

Although virtually all platforms are able to perform well in the baselines test, the i845 takes about 40% longer to complete the script than both the i850 and P4X266.  The two big bandwidth platforms offer the same performance levels once again.

Increasing the load does nothing to the standings, the P4X266 and i850 are still performing the same.

The i845 becomes very disappointing here.  The i850 and P4X266 are able to complete this test in almost half the time. 

In spite of our original assumptions that we'd see a difference between the i850 and P4X266 here, we saw nothing of the sort.  We're half-way through our benchmarks and so far it seems like VIA has a winner but let's save conclusions for later.

3D Rendering & Animation Performance

Here we have two tests: a 3D Studio MAX rendering test and the SPECviewperf benchmark suite.  The 3D Studio MAX rendering test stresses the CPU and memory subsystems while the SPECviewperf suite stresses those two areas as well as the graphics card (in this case, a GeForce3). 

It's obvious that our test scene was not large/complex enough to require more storage outside of the Pentium 4's 256KB L2 cache.  The result of this is that all three platforms perform about the same, in spite of vastly different memory bandwidth capabilities. 

Things aren't so simple in the SPECviewperf tests, as the sample sets are large enough to be memory bandwidth bound.  Some tests such as the ProCDRS-03 viewset are limited by the GeForce3 GPU, while others such as AWadvs-04 are influenced tremendously by memory bandwidth.

3D Gaming Performance

The final area of performance measurement is in a few 3D gaming tests using present day engines as well as future engines.

Quake III Arena has always been a strongpoint for the Pentium 4, and we may have just found out why.  This is the only test in which the i850 exerts a large lead over the P4X266, in this case a 16% lead.  The added bandwidth could very well be used by the Pentium 4's hardware prefetch indicating that a large portion of why the Pentium 4 does so well in Quake III Arena is because of its hardware prefetch. 

At higher resolutions this performance delta will shrink a bit as the system becomes more bottlenecked by your video card; it's simply interesting to point out.

Under DroneZ we see another somewhat large performance advantage, this time of around 8%. 

AquaMark is a good example of what happens in a benchmark when performance is limited by the graphics card.  All of the performance deltas disappear in the AquaMark test, leaving the P4X266 and i850 on level ground about 18% higher than where the i845 stands.

Final Words

When RDRAM was introduced as a solution for the Pentium III it wasn't much of a surprise that regular PC133 SDRAM could offer greater performance.  The Pentium III couldn't use the added bandwidth RDRAM offered and the latency penalty incurred was simply too great.  With the Pentium 4, the situation is much different.  The move to DDR SDRAM is a step down in terms of theoretical performance but luckily, thanks in part to VIA's excellent DDR controller, that doesn't translate into a loss of real world performance (in most cases). 

You cannot ignore the truth which is that the i850 with its controversial RDRAM does offer higher performance than VIA's P4X266.  The question isn't one about higher performance however; this time around the question is about, most bang for your buck.  It is in this category that the P4X266 was designed to excel.  VIA has never been a performance chipset manufacturer, they have always made chipsets for the masses and that's what the P4X266 is.

For the most part, the P4X266 is a viable alternative to the i850.  It is only under games that you lose a significant amount of performance and even then only at lower, non-video bottlenecked resolutions.  VIA's pricing on the chipset should be comparable to what they are selling the KT266 for, meaning that we should be able to see P4X266 based boards at or above KT266 price levels. 

Obviously the biggest attraction to the P4X266 is its ability to use DDR SDRAM which continues to enjoy a price advantage when compared to RDRAM.  As long as that price advantage lasts, it'll be much more desirable to go with a P4X266 than an i850 (motherboard costs aside).

Unfortunately the P4X266 will not be embraced by motherboard manufacturers right away.  VIA has a lot riding on the chipset, but then again so does Intel.  Most motherboard manufacturers have told us that they're simply going to keep quiet about it until Intel gives VIA the 'ok' to produce the P4X266.  While VIA is going to make the chipset regardless of what Intel may threaten, there isn't a single motherboard manufacturer in Taiwan that wants to be the first to ship a P4X266 board and actively promote it. 

Even Intel is divided over the P4X266.  There are some groups at Intel that are very much in favor of the release of the P4X266, mainly because of its ability to push for increased Pentium 4 sales.  There are others however that are adamantly in opposition to its entry into the marketplace and they will not stop at anything to make its birth as painful as it can be. 

The entire politics aside, what does it mean to you?  In the end, you get better choice in platforms and a lower cost of ownership for the Pentium 4.  With the transition being made to cheaper microPGA packaging and a smaller, cheaper to produce 0.13-micron core, next year could turn out to be a very promising one for the Pentium 4.  Remember that it took over a year for the original Pentium to actually become attractive…