Navigate: Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors WebNews Home	.	.
Prices: Mainboards ABIT ASUS Chaintech Shuttle Soyo Tyan CPU Intel P4 2.4C-800 P4 2.6C-800 P4 2.8C-800 P4 3.0-800 P4 3.2-800 AMD AthlonXP XP 1700+ XP 2000+ XP 2400+ XP 2500+ XP 2700+ XP 3000+ XP 3200+ Athlon64 Athlon64 3200+ Athlon64 FX-51 Opteron Opteron 240 Opteron 242 Opteron 244 Opteron 246 Memory Corsair Crucial Kingston Mushkin OCZ Search Prices:

In a side by side comparison of the 3.0 and the 3.2 GHz processors, both behaved very similar, however, it turned out that the "older" 3.0 GHz outperformed the new 3.2 GHz sample. The main difference was that the 3.0 GHz CPU manages to boot into WindowsXP at 3750 MHz and, furthermore, was running everything rock solid at 3.6 GHz whereas the 3.2 GHz sample showed the first signs of weakness at 3.5 GHz. We are not complaining, it is simply that our 3.0 GHz sample is a cherry processor, and the 3.2 GHz is probably more mainstream, that's all.

What is very clear from our data is that, first of all, performance has received quite a boost from moving from Granite Bay to Canterwood or Springdale. Moreover, the overall system configuration is still capable of pushing things a bit further without running into a stone wall. This is made possible by the high PSB in conjunction with the extra-wide memory bus (my personal feeling is that the term Dual Channel in this context is a misnomer since it is a single channel configuration).

In addition, the SATA RAID 0 interface does its part to alleviate the I/O bottleneck. It is actually quite interesting to look at the actual percentage of I/O accesses in the grand scheme of for example ContentCreation or Business Winstone. Since ContentCreation Winstone 2003 did show a very good scaling with increased clock speed, it appears as if HDD access plays only a minor role. We captured the I/O traces of CCWS2003 to get some idea of what is really happening and, surprise, overall disk access in a 512 MB system memory configuration is all but roughly 20 seconds (out of some 20 minutes total runtime). It is rather obvious that even if there is a bottleneck, it will be minor. Keep in mind, though, that the 20 seconds reflect the SATA RAID 0 setup and can easily grow into 2 minutes or more using slower storage devices. In that case, we would see the scores plummet.

One additional trend that surfaced was that there is still some sort of data congestion somewhere in the PSB. That is, as we showed in previous reviews, the overall memory bandwidth can be increased by increasing the PSB alone with the memory kept at constant speed. However, we also showed, that, despite the increased raw memory bandwidth measured by Stream and its derivatives such as SiSoft Sandra, Aida and others, performance suffers, courtesy of the increased latencies.

On a side note, it is interesting that increasing the memory frequency by running synchronously with the PSB does not increase the performance over the asynchronous setting, on the contrary, if the higher PSB has to be bought with higher latencies, performance will suffer. That is, unless streaming memory benchmarks are true indices for system performance but not even the flash and dance memory vendor community will twist the truth that far.

Back To The Future

The highest system performance was achieved at the 300 MHz PSB with the lowest multipliers and of course, the clock frequency was highest, too, in this case. However, given all that we have looked at today, it appears as if the 800 MHz PSB is not the end of it. Maybe for Northwood it will be but Prescott has the feet already in the start holes. Yet higher PSB interfaces together with advanced HT in combination with SATA native command queuing and a potential move to cached DRAM architecture are most likely the next Intel moves. After all, Elpida owns the IP for virtual channel DDR and Intel just invested heavily there. All three technologies, that is HT, NCQ and VCDDR (or EDDR) by themselves may be wallflowers, however, combined with each other, they will be dynamite since they are perfect complements for each other. Suffice it to say that not only with respect to performance but also concerning techno-politics, intellectual property and licensing fee wars, the next few months will turn more than a few rocks over.

next page:    => more =>