Well, DDR333 memory is little by little becoming an industrial standard. There are pretty many chipsets for different platforms in the market already, which support this memory type. JEDEC approved the spec and many memory makers started mass production of DDR333 memory modules. This way the bandwidth of the mass DDR SDRAM has finally increased up to 2.7GB/sec.

However, is there any real need for DDR333? As our tests showed, using faster DDR333 makes sense for Pentium 4 platforms, providing a noticeable performance increase. And that is not surprising. Pentium 4 processor bus bandwidth makes 3.2GB/sec, which means that systems like that have the chance to involve the entire bandwidth of DDR333 memory. As for Athlon based systems, the situation is a bit different here.

The system bus frequency there has stopped at 266MHz, which means that the processor bus bandwidth is only 2.1GB/sec. We have already seen what comes out of it. The performance grows just a very tiny bit when we shift from DDR266 to faster DDR333.

However, we would like to remind you that EV6 processor bus used in Athlon systems was initially designed to be able to work at up to 400MHz. so theoretically, no one can prevent AMD from speeding up the Athlon CPU bus another time, up to 333MHz, for instance. In this case, the systems based on processors with faster bus will boast a significant performance increase as more data will be transferred between the processor and the memory.

Nevertheless, AMD doesn't seem to be so willing to increase the CPU bus frequency. Athlon XP processors with 266MHz bus will stay with it later on. Even the upcoming shift to finer 0.13micron manufacturing technology and Thoroughbred core will not change this state of things. All Athlon XP CPUs coming out at least this year will not get a faster bus anyway.

Why doesn't AMD feel like making this evident step towards easy performance boost? This question is especially acute since there are a lot of DDR333 chipsets for Socket A systems in the market already, which could use fast processor bus and fast memory bus to the system advantage. Take, for example, VIA KT333, which can combine 333MHz processor bus with DDR333 memory, so that the CPU and memory buses could work synchronously. It could help to avoid the incompatibility of their peak bandwidths. Moreover, increasing the system bus frequency seems to be a more evident means of raising the performance of an Athlon based system than even the increase in the L2 cache size. However, AMD stakes L2 cache size as the No 1 solution. In H2 2002 its size will grow up to 512KB while the system bus frequency will remain equal to only 266MHz.

Well, now it's high time we passed over to answers to all these questions mentioned above. Of course, it is not for nothing that AMD doesn't want to use 333MHz CPU bus. And these reasons have to do not only with marketing, but also with engineering.

First of all, you should bear in mind the fact that AMD is already looking forward to the upcoming launching of the ClawHammer processor family in the end of the year, and Athlon XP is of no priority any longer for the company. Moreover, the announcement of the ClawHammer baby positioned as a performance solution will oust Athlon XP to the value market now occupied by Duron and Celeron CPUs. Therefore, AMD started "holding back" Athlon XP family, so that to push it smoothly and painlessly to a new position. Besides, a very aggressive advertising and promotion campaign will definitely accompany ClawHammer launching, thus the more advantages ClawHammer will be able to present over Athlon XP, the better. So, AMD will do its best to make the ClawHammer bus bandwidth much higher than that of the Athlon XP CPU. This was the marketing reason.

Also, a shift to 333MHz bus (or 166MHz FSB frequency) will lead to some engineering problems. Although Athlon XP processors are already capable of working with faster bus, we wouldn't say the same about mainboards. The increase in the FSB frequency will lead to greater EMI, which in its turn will cause overall system instability. That is why the implementation of faster 333MHz bus will require some extra tests to be carried out, and maybe even a complete layout redesign. Far not all the mainboard makers are ready for extra expenses. That is why no wonder that only a half of all Socket A mainboards available in the today's market can work well with 166MHz FSB.

Nevertheless, we wouldn't deny the possibility of implementing 333MHz processor bus in Athlon XP CPUs. For example, the recently announced DDR333 chipset from VIA aka KT333 does support 166MHz FSB, although no specs sheet declares that. Here it means that once the FSB gets equal to 166Mhz the PCI and AGP buses still work at the nominal frequencies. In other words, the chipset allows setting the FSB-to-PCI frequencies ratio as 1:5. Certainly, all the mainboards based on VIA KT333 chipset support the same option. Note that even with 166MHz FSB, the memory frequency will remain only 166MHz because of some limitations imposed by the chipset. However, it is exactly this particular case (166MHz FSB and DDR333 memory) that arouses most interest by experiment-lovers.

Here it should be also mentioned that setting 166MHz FSB frequency is no big deal for any VIA KT333 based mainboard (the AGP and PCI bus frequencies will be equal to 33MHz and 66MHz respectively). However, far not all the boards will work stably with these settings. Many mainboards on VIA KT333 owe their instability to older PCB design borrowed from the VIA KT266(A) based products. Moreover, many mainboard makers regard 166Mhz FSB as an overclocking friendly feature and hence to not run any stability tests for their products working in this mode. All in all, we could recommend only two products out of four VIA KT333 based mainboards tested recently (see our article called: "5 Socket A Mainboards Supporting DDR333 on VIA KT333 and SiS745"). They are: ASUS A7V333 and MSI KT3 Ultra. Another couple of mainboard didn't show any stable operation at 166MHz FSB frequency.

Well, enough of talking, let's pass over to some practice.

Testbed and Methods

This test session was aimed at finding out if the use of 333MHz processor bus is justified for Athlon XP processor family. We tried to see how greatly the Athlon XP performance grew due to the use of faster system bus. So, for our investigation we took Athlon XP 2000+ CPU (working at the actual 1666MHz). We unlocked the clock frequency multiplier by closing 5 L1 Golden Bridges on the CPU surface. Then we measured its performance in three modes:

• 133MHz FSB and 133MHz DDR memory;
• 133MHz FSB and 166MHz DDR memory;
• 166MHz FSB and 166MHz DDR memory.

As a test platform we took one of the fastest VIA KT333 based mainboards - MSI KT3 Ultra-ARU. We would like to stress that even when the FSB frequency was set to 166MHz, this board proved highly stable throughout the whole testing session.

For a better and more illustrative comparison we also included the results for Intel Pentium 4 2.4GHz working in an i845D based system with PC2100 DDR SDRAM.

Here are our testbeds used:

All systems worked in Microsoft Windows XP.

Performance

To begin with, we suggest looking how greatly the memory subsystem performance depends on the system bus frequency. Here are the results of the synthetic Cachemem benchmark obtained by taking the memory subsystem bandwidth when reading, writing and copying large data packs (larger than the CPU L2 cache).

The results show that the real memory subsystem performance is limited by the processor bus bandwidth only during reads. The faster system bus in case the memory frequency remains the same provides a certain effect on the performance during write and copy operations, although this effect is not that big. This way, DDR333 used in Athlon based systems ensures faster writing and copying into the memory. As for reading from the memory, it cannot be speeded up in any way because when the data is transferred between the CPU and the memory it has to pass through a bottleneck between the CPU and the chipset. Raising the FSB frequency up to 166MHz will help eliminate this bottleneck and improve the data transfer rate.

We also noticed one very curious thing. Pentium 4 processor working with DDR266 memory reads and writes into the memory faster than Athlon CPU with faster memory and 166MHz processor bus. It is probably the Palomino core that is to blame here, as it reads from the memory slower than the good old Thunderbird. Even VIA KT333 combined with an Athlon on Palomino core appeared unable to use the entire potential of DDR SDRAM. So, AMD and VIA engineers still do have something to work on.

The latency measurements look even more exciting. As we can see, the processor bus frequency growth in Athlon systems causes the latency CPU requests sent to the memory subsystem to get lower. However, you should bear in mind that the memory latency itself doesn't get any lower in this case. If we divide the obtained results by the clock frequency multipliers of the CPUs tested, we will get 366/24=15 clocks for i845; 196/10=19 clocks for VIA KT333 and 166MHz FSB; 227/12.5=18 clocks for VIA KT333 and 133MHz FSB.

I would also like to draw your attention to the fact that i845D, which has already pleased us in the previous test, again proved better than VIA KT333. And again we have to complain about the memory controller implementation by VIA chipsets.

The results obtained by the memory subsystem in SiSoft Sandra 2002 benchmark prove our conclusions once again. DDR333 turns out much more efficient if not only the DDR SDRAM frequency grows but also the processor bus frequency. In standard VIA KT333 mode, however, that is when the FSB frequency is 133MHz, the use of faster DDR333 memory instead of DDR266 hardly tells on the performance. This observation lets us conclude that it would make much more sense to use DDR333 in Athlon systems supporting synchronously working 166MHz CPU bus. But the fate decreed differently…

Now we pass over from the synthetic tests to real applications:

For office and content creation applications, we used a new SYSmark2002 test set. It differs from the previous benchmarks set dating back to 2001 only by the versions of the included applications. Nevertheless, the memory bus got loaded twice as heavily compared with what we had in case of SYSmark2001. Now the updated test package is even more sensitive to the memory subsystem performance.

The results show that with DDR333 the system performance grows by about 2%, and the use of faster system bus adds another 3% of performance. Also, note that faster memory doesn't have any positive effect without increasing the processor bus frequency in internet content creation applications, while in office applications both: the memory and the processor bus contribute to the overall performance growth.

As we see, archiving speed depends a lot on the memory subsystem performance. However, using DDR333 together with asynchronous 133MHz processor bus won't be that efficient. However, when you increase both: the FSB and memory frequency from 133MHz up to 166MHz, WinRAR algorithm will start working 16% faster.

We can't call mp3 encoding an operation depending too much on the memory performance. Look here:

The same thing is valid for video compression. The whole thing doesn't get any faster even when the working frequencies of the CPU bus and memory grow up. It looks as if 2.1GB/sec memory bandwidth were more than enough for audio and video encoding algorithms.

3DMark2001 SE test modeling the future games shows that DDR333 memory fails to affect the performance in case the processor supports 133MHz bus. But as soon as the FSB frequency gets equal to 166MHz, the performance improves by 4%.

Well, and this is the secret of Pentium 4 success in Quake3 Arena. Everything depends on the memory subsystem performance. If AMD could provide its CPUs with the 333MHz bus, they would outpace Pentium 4 in a wink here.

In the recently released version of a popular helicopter flight simulator, the data transfer rate between the processor and the memory appears the key factor. Just replacing DDR266 with faster DDR333 will not change anything.

The diagram shows another argument in favour of 166MHz bus, though it doesn't seem to matter for AMD at all.

3D designers may also benefit from the Athlon CPU bus frequency increase. By making the FSB 25% faster (with DDR333 memory used) you can get almost 8% of extra performance in ViewPorts.

The results obtained in the professional SPECviewperf test prove the mentioned above statements.

Well, all this evidence seems to be pretty convincing, eh?

Conclusion

So, we saw that the performance can be improved significantly by simply increasing the bus frequency of Athlon XP processors. When we switch from 133MHz FSB to 166MHz FSB (with DDR333 memory used) the performance gain makes 7%-15%. It's a pity that AMD will never take advantage of this opportunity. And it could be of real help especially taking into account that the competition between AMD and Intel has got even tenser after Intel had introduced 533MHz bus.

However, no one can prevent hardware enthusiasts from getting their FSB to work at 166MHz, even without AMD's assistance. In this case they will need DDR333 memory, high-quality mainboard based on VIA KT333 and AMD Athlon XP processors with an unlocked clock frequency multiplier.