Architecture continued

So why would people even consider the NF3 150 chipset and why would it become so popular? Well, while there was a slight disadvantage to using a board with a slower overall bus, the NF3 150 did offer one feature the Via KT800 did not, a locked AGP bus. While the memory controller was integrated onto the CPU itself, the motherboard still had the responsibility for controlling the AGP and PCI buses. As HTT scales, an unlocked AGP and PCI bus will eventually rise to the point at which certain components become unstable due to improper frequencies. ATI Radeon 9700/9800 cards were notorious for not liking HTT buses much past 220-230 MHz, which coincidentally was about the point to which SATA hard drives began to corrupt data. The NVIDIA boards using a locked AGP bus didn't suffer from this problem. Many also believed that the NVIDIA boards featured locked PCI buses as well, given that several users were able to run 300+HTT speeds without drive failure, which seems logical as I don't know of any hard drives that would tolerate the 50 MHz PCI bus that would occur if it was unlocked.

Second generation Athlon64 boards improved on prior faults and NVIDIA NF3 250 boards could run a 5x HT multiplier or 1000 MHz Hypertransport bus or 2000 MHz effective, as could VIA KT800Pro chipset motherboards. The NVIDIA NF3 250 featured advertised working AGP and PCI locks. The Via KT800pro was supposed to have these features, but initial revisions of boards were problematic to say the least in regards to the supposed AGP and PCI locks.

Yet another variable on system performance was the ram multiplier or more accurately what speed the CPU interacted with system memory. Since the Athlon64 had no traditional FSB, there is really no true synchronized ram speed, which implies the memory runs at same frequency as fsb. This would be the case in Intel 865/875 motherboards whereas memory running at 1:1 ratio would run at the frequency of the front side bus and at default levels 200 MHz bus and 200 MHz memory (400 MHz effective due to DDR). In the case of the A64 the memory controller(and hence the fsb) is moved onto the CPU die, so is in effect running at full core speed. Since we don’t have memory that can run at the frequency of the A64 cpus, there is no true synchronized ram speed to the bus and any setting is asynchronized to the CPU speed.

Instead you may choose what you wish your ram to run at in relation to the HTT. For example, you may select DDR400/PC3200, DDR333/PC2700, and DDR266/PC2100. Some board manufactures may also allow other memory speed settings, but these mentioned are by far the most popular. To make matters even more complex is that while I just stated there is no true "synch" ram divider, guess what we enthusiasts use as the term to identify what ram running inline with HTT bus? Yep, "synch"! While this is technically wrong. “Old dogs and new tricks”, if you know what I mean!

People confuse this argument all the time, so don't be concerned if you are confused at this point. What I would like you to get out of this history and platform explanation is the how the chip operates and interacts with the motherboard. I also feel that you should not become caught up in the idea that you need to maintain “synch” or DDR400 ram speed to get good performance from an A64 system. I will grant that optimally you run you HTT and ram as high as you can by forcing DDR400 speed to the ram. When overclocking the HTT you will force the ram to accompany the HTT at whatever speed the HTT is running. However, if you do not have ram that can scale as high as your HTT bus can, don't be afraid to use a lower ram speed in an effort to scale your HTT higher if your concern is ramping cpu MHz. We will get back to this idea later in an effort to explain why running a lower ram speed is not disastrous to performance. The benefits of the on die memory controller, and the HTT bus serve to decrease performance losses in running lower ram speeds. Additionally, the architecture of the A64 is such that the system is not penalized severely for lower ram speed/MHz as much as it’s competitor in the Pentium4. One should consider memory latentcy more important that memory speed in general for the A64. Now there will come a point where higher ram speeds will trump low speed and latentcy, but for the sake of this argument, we can still accept good performance from using slower ram speeds at tight latency.