Hey Sarge, OCZ is good stuff from what I have heard, and Corsair do some very nice memory too.
Here's the thing about RAM, there are two distinct variables which control potential performance:
1) Rated DDR speed (in MHz or PC rating)
2) Latency ratings, typically expressed like 2-5-5-7
The DDR speed (i.e. DDR400) relates to TWICE the Front Side Bus Speed. So, DDR400 is rated for operation with a 200MHz FSB. DDR500 = 250MHz FSB.
The PC rating is simply 8 times the MHz rating and reflects the number of MB per second which is theoretically available. It's 8 times because it's 8 Bytes per clock transfered across the bus.
So, DDR400 = PC3200 (3.2GB/s bandwidth). With Dual Channel RAM, that increases to 6.4GB/s because each bank of RAM has it's own bus.
Overclockers will typically want RAM rated at significantly faster than there "stock" FSB speed, because raising the FSB is the only way (usually) to overclock a CPU.
See, a 3.2 GHz P4 uses a 200 MHz FSB (Yes, it *says* 800MHz, but that is "quad pumped") with a CPU multiplier of 16 to arrive at 3.2 GHz. (200 x 16 = 3200)
If your system is 200MHz FSB, and you install DDR400 (which is specified) then you may not have much luck overclocking, because you'd be pushing the FSB beyond the rated speed of the RAM.
If you installed DDR500 (PC4000) then you will KNOW that you can push the FSB all the way to 250 and the RAM will still be inside the spec. (Other peripherals, or your CPU might stop you getting that high though.
For example, a 3.2GHz P4 using 250 FSB (1000MHz quad-pumped) will actually run at 250 x 16 = 4000 MHz. It would be unlikely you'd get a 3.2 GHz CPU to 4.0 GHz without extreme cooling though.
So - if you intend overclocking, get a speed rating higher than the rated one. I spec DDR433 (PC3500) and DDR466 (PC3700) on machines which will be overclocked.
Item 2 is the latency - and you pay through the nose for low latency RAM. Low Latency simply means you wait less clocks before data streams from the RAM. Random Accesses into memory are delayed by the latency ratings of the RAM.
Conventional wisdom says the speed increases available with RAM that costs twice the price of budget RAM (at the same DDR rating!) are not worth the expense - unless you have specific requirements for low latency RAM.
So, it's better to buy high DDR ratings, as opposed to Low Latency RAM at the correctly rated DDR speed.
It gets more complicated though, because if you buy budget fast RAM (i.e. cheap DDR466) but only run the RAM at 215 MHz (DDR430 speed) then you MAY be able to switch away from SPD settings for RAM in the BIOS, and then try to lower the latency settings a little.
Conversely, if you buy very low latency DDR400 but push it to DDR433 speeds, it might fail unless you increase the latency settings manually.
================================
Low Latency = more expensive than higher speed.
================================
There is an issue with the Pentium 4 though, Prescott included. WITHOUT Low Latency RAM, the CPU can sit there spinning it's wheels waiting for data from memory. Because of the long CPU pipeline in the P4,a huge number of clocks go wasted - so to keep a busy P4 core happy, you have to not only provide it with huge amounts of data, but deliver it VERY soon after the request for the data. A huge hole in two CPU pipelines is just guaranteed to slow you down a lot.
This is one of the big reasons why AMD at 33% less outright core clocks, can outperform a P4 in many tasks: AMD have a much shorter pipeline, so holes don't hurt you as much, and the on-die memory controller offers far lower latencies with "standard" RAM.
Another reason the P4 is easy to beat in gaming, is because the P4 uses a LOT of clocks to do branch prediction: it predicts what the outcome will be and does a lot of calcs for each possible outcome. When the probability tree collapses to a single option, the P4 has wasted a lot of time if it didn't choose a correct branch to follow. (This is the reason P4 excells at media encoding: branch prediction almost never fails!)
Therefore, the AMD architecture is more forgiving about RAM. Note: the socket 754 AMD64s have half the memory bandwidth (3.2 GB/s = single channel) of the Socket 939 cores, and the P4. Note also: The socket 754 Athlons perform almost up to the level of the Dual Channel architectures!
That was probably more than you wanted to know though ...