Double Data Rate SDRAM was the talk of the town by the end of 2000, and now that it is here, the market can't find enough ways to put it to use. There is a tremendous backlash against the business practices of Rambus, the chief proprietor of DDR's closest competitor, and it seems as if a lot of the support for DDR grew out of disgust for Rambus.

In synthetic performance tests and forward looking benchmark comparisons such as those we have performed under SPEC CPU2000, DDR SDRAM definitely has a bright future. However in terms of offering tangible performance benefits today, the only advantage DDR SDRAM offers over RDRAM is that it isn't any slower than PC133 SDRAM in everyday applications and games.

There is no doubt about it though, as we do begin to see newer applications and games hit the market, there will be a much greater demand for a higher bandwidth memory solution. With this in mind, the scene is now set for DDR technology to have a banner year in 2001. The industry has been asking for DDR technology to be brought into the system memory market and now that it is here there is the question of how to promote it.

Currently the price gap that exists between DDR SDRAM and PC133 SDRAM is no where near the 8 - 10x levels that PC800 RDRAM was at just a year ago, but that doesn't mean that the price gap is acceptable. Companies like Crucial, in an attempt to gain some control in the market, have already begun offering PC1600 DDR SDRAM at price points identical to what you can find PC133 SDRAM at. Twice the bandwidth at the same cost, a very good marketing slogan if you ask us.

Where there is memory, there are also platforms to take advantage of it. AMD kicked off the DDR bandwagon late last year with the release of their 760 chipset. A combination of its 133MHz DDR FSB and PC2100 DDR SDRAM support resulted in a 10 - 15% performance gain in present day applications and games. There were definitely cases where the performance improvements grew even beyond those figures, but for today's user, the performance improvement was generally within that range.

A somewhat disappointing ALi release followed the AMD 760, and although the MAGiK1 chipset has done some maturing since we first took a look at it, it is still in its relative infancy. Interestingly enough, throughout the end of 2000, VIA was missing from the DDR scene. Although they officially announced their first DDR chipset in September of 2000, we did not see any motherboards based on the Apollo Pro 266 back then.

Instead of promoting DDR platforms, VIA wisely chose to focus on what was important to them in the market, producing and shipping platforms that were currently in great demand. This brought the introduction of the KT133A chipset, which although won't offer the security of high future performance, it does present the best price/performance match for the Athlon right now. We also saw the long awaited release of the KM133 chipset from VIA as well, the first value PC solution directed at the Duron market from the manufacturer that had supported the Socket-A platform so well since its introduction last June.

While VIA's Socket-A platforms might be the most talked about, their Socket-370 chipsets are what paved the way for them to gain the market share they currently enjoy. It wasn't surprising, then, that VIA's first DDR platform would be directed at the Socket-370 market because at the end of the day, unlike many of the dot-coms of recent history, VIA's intent was to produce a profit. This isn't to say that the Socket-A platform isn't profitable, but VIA felt that pursuing the Socket-370 market first made the most sense since they are potentially dealing with a larger volume of sales; not to mention that VIA pretty much had no competition in the Socket-A market to worry about.

With that said, in September of 2000 came VIA's announcement of their first DDR SDRAM capable chipset: the Apollo Pro 266.

Haven't we been through this before?

The whole premise of our VIA KT133A review was to illustrate that the Athlon, in its current form, was not a memory bandwidth hungry processor leading us to believe that DDR SDRAM on the Athlon would not have the same effect as RDRAM did on the Pentium 4. If the Athlon isn't a memory-bandwidth hungry CPU, then the Pentium III is definitely much less of one.

Remember that the Pentium III was released with the assumption that people would be using PC100 SDRAM, and we saw the non-existent performance gain that pairing it up with RDRAM produced. In order to prove this theory, let's head back to our good friend Linpack and take a look at what happens when we compare performance between a Pentium III on a platform with 1.06GB/s of peak available memory bandwidth and one with twice that, 2.1GB/s.

As the data set gets larger than the Pentium III's L2 cache we can begin to measure the performance of its memory bus. Just like the Athlon, the Pentium III's performance does not skyrocket when paired with a memory solution that offers twice as much bandwidth.

A good comparison point here is the Pentium 4, which is based on an architecture that is so very memory bandwidth dependent that giving it 25% less memory bandwidth results in a 17% decrease in floating point performance, according to Linpack.

Unlike the Athlon however, the Pentium III is nearing the end of its lifespan. The only remaining update (Tualatin) to the current Pentium III core will revolve around giving it a lower power consumption figure and a potential for higher clock speeds courtesy, primarily, of a new fabrication process and a resulting die shrink.

The Athlon's outlook is much more positive. AMD has two more core updates on their current Athlon roadmap, the Palomino and the Thoroughbred. Although the Palomino may not gain the core improvements that we'd like to see outside of reducing power consumption, it would be a sin for AMD not to implement a few new features into the Thoroughbred core.

The bottom line is this: just as the Athlon did not respond to DDR SDRAM with astounding performance improvements, the Pentium III will most likely offer much less of a response.

Then why produce it? The potential for DDR SDRAM to be offered at a price point identical to that of PC133 SDRAM is much greater than the ability for RDRAM to do the same. In fact, we are fairly confident that DDR SDRAM will be presented in the very near future as a solution offered at prices that consumers were previously paying for PC133 SDRAM. With that in mind, combined with the marketability DDR in general, VIA would be insane not to produce as many DDR capable chipsets for as many different platforms as possible.

The Apollo Pro 266

As opposed to the past few chipset releases from VIA, the Apollo Pro 266 is a significant departure from the Apollo Pro 133A If you remember, the KX133 and KT133 chipsets were actually not so distant relatives of the Apollo Pro 133A, they shared the same memory and AGP controllers and with VIA's flexible designs they used the same South Bridge controller as well. The only real difference between the Apollo Pro 133A North Bridge and that used on the KX133 and KT133 chipsets was that the latter interfaced with the Athlon's EV6 bus.

The one thing that has remained constant throughout every single VIA chipset release, even dating back to the days of the old Socket-7 chipsets is the fact that the PCI bus has always connected the North Bridge and South Bridge. That's right, the 32-bit, 33MHz Peripheral Component Interconnect bus is what has "bridged" the gap between the two major chips in all VIA chipsets. This was true of all Intel chipsets up to and including the i440BX before they moved to the Intel Hub Architecture.

The "old" way: PCI Bus connects North & South Bridges

As a quick refresher, the North Bridge is what houses the memory and AGP controllers. It is also the chip that lies connects directly to the CPU, thus controlling your FSB frequency among other things related to your CPU bus interface.

The South Bridge handles any ISA bridges, the integrated hard drive controller, it contains the USB controller and in the case of VIA's 686A, it also houses the keyboard/mouse controllers.

In a conventional chipset such as the Apollo Pro 133A and the i440BX, the North Bridge is connected to the South Bridge via the PCI bus. As we mentioned before, this is a 32-bit connection that operates at 33MHz resulting in 133MB/s of available bandwidth. All of the PCI devices, as well as any traffic that occurs between the North and South bridges share this bandwidth.

With current IDE hard drives approaching 40MB/s sustained transfer rates, USB devices becoming increasingly more popular and 100Mbit Ethernet cards eating up this 133MB/s of available bandwidth there is a need to remove this potential bottleneck from the system.

VIA's Solution: V-Link

Intel identified and addressed this issue back in 1999 with the release of the i810 chipset. The i810 was the first to employ what is now known as the Intel Hub Architecture. To make a long story short, IHA replaces the PCI bus as the connection between the two main parts of the chipset. A bus capable of achieving twice the available bandwidth of the PCI bus replaces it. This bus is an 8-bit wide bus that operates at 133MHz DDR, effectively resulting in 266MB/s of available bandwidth. This interlink bus is present on all 8xx chipsets, including the popular i815.

The "new" way: A dedicated bus connects North & South Bridges

VIA's solution to the problem, albeit a couple of years later than Intel's Hub Architecture, is very similar. Dubbed V-Link, the connection between the North Bridge and the South Bridge is now this interface capable of transfer rates of up to 266MB/s, identical to Intel's interlinks bus.

On the Apollo Pro 266 Reference board you can actually see the traces that make up the V-Link bus quite clearly:

Click to Enlarge

Without a need for the PCI bus to connect the two chips it is now moved down to the South Bridge along with the ATA/100, USB and integrated 10/100 Ethernet controllers.

The Apollo Pro 266 in particular is composed of the VT8633 V-Link Host North Bridge and the VT8233 V-Link Client South Bridge. The South Bridge is virtually identical to the 686B from VIA with the exception that it now contains the PCI bus interface, an integrated 10/100 Ethernet controller and supports up to 6 USB ports.

The Test

The board we used to test was the Chaintech 6VJD as pictured below:

Click to Enlarge

The blue slot that looks like a backwards PCI slot is actually an ACR slot.

Business Winstone 2001 is a good example of the limitations of disk I/O performance in today's systems. While the 2001 version of the business level benchmark is bottlenecked much less than the older versions were, it is still a factor. After that, obviously L2 cache and memory performance come into play but since we aren't dealing with any extremely bandwidth intensive benchmarks DDR SDRAM isn't really given an opportunity to shine.

In this case, the Apollo Pro 266 is penalized twice. First, as we just mentioned, because the benchmark isn't very memory bandwidth intensive and second because of the inability of the Pentium III processor to actually take much advantage of the added memory bandwidth.

The performance the Apollo Pro 266 is capable of offering is very close to that of the i815E which does have a much more mature memory controller as well as more mature motherboard designs to support it.

Content Creation Winstone 2001 is more of a memory bandwidth hog than its business performance counterpart, however it is still far from demanding the 2.1GB/s of available memory bandwidth the Pro 266 can offer. This unfortunately will be the case for quite a few of today's users since this particular benchmark does test performance in a manner similar to the way many use their systems. For more information on the tests run, have a look at our quick description here.

Quake III Arena has consistently shown us that it not only responds with a decent performance increase when given more FSB bandwidth but also when given more memory bandwidth. In this case, the Apollo Pro 266 is able to almost match the i815 in terms of performance. While this may not seem like much of an accomplishment you'll have to remember that the i815 has actually been a very difficult competitor to beat in this test.

This doesn't mean that you should throw away all of your PC133 SDRAM and pick up a Pro 266 board, however looking towards the future, if the motherboard and DDR SDRAM is priced identically to an 815 + PC133 SDRAM combo then there will be very little to discourage going after the Apollo Pro 266.

At 1024 x 768 x 32, as we have proven countless times, a game like Quake III Arena is limited by the memory bandwidth available not to our CPU but to our video card. This results in the nearly identical performance across the board.

With the next-generation of graphics cards due out in a few months now this limitation should hopefully be addressed soon enough.

MDK2 is very reminiscent of Quake III Arena in the type of stress it leaves a system with. The Apollo Pro 266 is able to pull close to the i815 however it actually falls further behind here than it did in the Q3A tests.

Again we see that the GeForce2 GTS test video card is memory bandwidth limited at 1024 x 768 x 32.

UnrealTournament has historically been more of a memory performance test than Quake III or MDK2, however in this case the Pro 266 still falls approximately 2% behind the i815. A more mature Pro 266 board may be able to narrow that gap even further, and possibly even tilt the balance in favor of the Pro 266 chipset.

We get a very familiar picture as the resolution increases however there is more of a performance range here because of other limitations acting on the overall system performance under UT.

To conclude our gaming performance tests we have Expendable which is more easily influenced by memory bandwidth than many of the previous benchmarks. Going along with this assumption about the benchmark it isn't too surprising that the Apollo Pro 266 is able to slightly edge out the i815 for the first time thus far.

Under SPECviewperf, which is designed to be a simulation of system usage as a professional level 3D workstation, the Apollo Pro 266 actually offers performance that is either equal to or slightly greater than the i815. Again, if the price of DDR SDRAM were to come down to the level of PC133 SDRAM, the Pro 266 wouldn't be such a bad alternative to the i815.

Of course no DDR chipset review would be complete without the obligatory synthetic memory performance benchmarks courtesy of SiSoft Sandra's STREAM tests.

The memory bandwidth offered by the Apollo Pro 266 puts it exactly in between the i820 and the dual channel i840, making its score of 410MB/s here in line with the theoretical numbers. It is interesting to note how much higher all of the Athlon platforms perform here by default, even with regular PC133 SDRAM.

A similar story is told in the FPU-STREAM test, the Pro 266 offers more theoretical bandwidth than the i815 and less than the i840 but we already know how that stacks up in most real world performance tasks.

Final Words

The Pentium III isn't going to be able to take much advantage of DDR SDRAM, much like it wasn't able to take advantage of the higher bandwidth RDRAM offered when it was first introduced to the P3 back in 1999. However unlike the i820/RDRAM launch of 1999, VIA may have hit on something interesting with the Apollo Pro 266.

Granted, there is no reason to upgrade your i815 to an Apollo Pro 266 unless you are absolutely in dire need of more memory bandwidth. Applications in which memory bandwidth is in great demand will obviously benefit from the Apollo Pro 266, however for the majority of AnandTech readers, the Pro 266 does not offer any tangible performance gains over the i815.

But that isn't necessarily what VIA hoped to accomplish with the Apollo Pro 266. If DDR SDRAM can get to a point where it is just as affordable as PC133 SDRAM, VIA may be able to package the Apollo Pro 266 + PC2100 DDR SDRAM solution as a healthy alternative to Intel's 815. It seems like the market wouldn't have much of a problem adopting a higher bandwidth platform if it were priced identically to the i815 + PC133 setup which is selling quite well today.

Being able to market a product and it being worth it are two different things however. And in this case, the Apollo Pro 266 is simply not necessary for the Pentium III platform from a performance standpoint. The introduction of VIA's DDR SDRAM controller and V-Link are quite possibly the two most important features that we should take from the Apollo Pro 266 chipset, since those two features will definitely find their way into other, more memory bandwidth hungry platforms.