SDRAM History 101:

When evaluating memory we're not strictly evaluating the name stamped on the heat spreader or package, but what's under the hood as well. With the exception of a handful of companies one of the largest being Samsung another somewhat smaller being A-Data most memory supplier's outsource IC's and other parts such as the PCB. Initially misunderstood prior to the introduction of the heat spreader, as PC users became more advanced, and overclocking became more popular, specialty memory supplier's such as OCZ, Corsair, Kingston, and Mushkin (to name a few) catered to a rapidly growing PC-enthusiast/Overclocking segment.

There were even scandals such as OCZ being accused of "...not making their own memory chips..." which gave early forum readers a soap-operatic/outlet to displace anger towards inferior products they'd purchased in the past. Of course the argument was at once whimsical and frustrating when trying to convince people practically none of the "specialty" memory companies actually owned Fab's, or made IC's. As End-users familiarized themselves with the manufacturing process they came to respect the effort it took to coordinate and oversee several manufacturing stages scattered around the globe. More successful companies found sources based on their perspective manufacturing talents, bringing the final product in-house where it was tweaked, tested, binned, and eventually shipped baring the company name. Alleged inside photo's of these "sweat shops" located along strip malls in the Colorado Rockies, became infamous on the MadOnion forums.

As the PC-enthusiast/overclocking community became an avant-garde force to be reckoned with, they sought to crack the memory manufacturing code. Overclockers discovered what' was under the heat spreader quite often determined performance regardless of the company name on the package. For example when it was rumored a specific IC was responsible for some of the fastest DDR on the planet, Overclockers left no stone un-turned and no heat spreader in tact to find it. That specific IC became the hottest commodity in the history of DDR memory, as it was the common denominator among the DDR speed demons. The code they cracked consisted of two letter's one number and a company name, and it changed everything BH-5, Winbond.



Ancient History:
That was then, this is now, and as we've begun to scale the DDR2 mountain-range the question on many parched lips is; "What will the summit look like?" Unfortunately we haven't established the first base-camp yet so the answer is far off. Motherboard makers and memory suppliers need to coordinate getting stable BIOS versions out. At this point I don't know of a single motherboard which can run DDR2 in any mode other then 533 or Auto, nor can the PCI-express bus be locked (if you know of one please write in).

Companies are working to tighten latencies on their DDR2 series, and Mushkin is continuing along the path they chose with their Black LVII series. This high-performance memory became popular with Overclocker's and PC-enthusiasts alike in a very short time. Economically friendly and faster as the fastest, Black LVL I and LVII were a more then a phenomenon.

Today while evaluating Mushkin's PC2-5300

Mushkin 1GB PC2-5300 Dual Pack
Speed	DDR2 667MHz
Pin Count	240 pins
Buffering	Unbuffered
Error Correction	non-ECC
Voltage	1.8V ~ 2.0V
Latency	CL4-4-4-11
Warranty	Lifetime

While the common denominator throughout many phases in DDR evolution had to do a lot with IC’s such as the infamous Winbond BH-5, it is not always the case IC's alone will dictate performance. This is especially true when the IC's in question have been squeezed by manufacturers to get every last MHz. The best companies produce some of the best performing memory based on engineering prowess. The ability to take mediocre IC's and run these well beyond their rated speeds while keeping latencies tight is a subtle balancing act. This is where companies such as Corsair, Kingston, and Mushkin whom may purchase the same IC's from the same company and even from the same batch must rely on good ole fashioned ingenuity to surpass the competition. With DDR2 latencies as they are this task may seem insurmountable. How ever will these companies drop latencies from 5-4-4-15 or 4-4-4-11 to something reasonable given the speeds? Memory suppliers have their work cut out for them, as do we.

Test Setup
CPU	Intel P4 550 3.4GHz LGA 775
Cooling	Alphacool Xtreme Pro Set with NexXxos BOLD
Mainboard	Abit AA8 Duramax (BIOS 17)
Graphics	ATI Radeon X800 XT PCI-ex
Memory	Mushkin PC2-5300 2x512MB DIMM's 1,3
Storage	Maxtor DiamondMax Plus 9 SATA150/120GB HDD
Power Supply	TTGI Modular 600W 4-fan BTX
Case	Thermaltake Shark Aluminum Tower
Software/OS	WinXP SP2, CPUZ, CPUID, Sandra, Aida32, Cahemem, 3DMark2001 (330), SuperPi, ScienceMark 2.0

The modules were testing today run at 667MHz DDR2 and similar to DDR as speed increases so too do latencies. The statement below is one of the clearest I've read pertaining to latency, it's from Ars Technica RAM Guide Part.II; The CAS latency is the delay, in clock cycles, between the registration of a READ command and the availability of the first piece of output data. The latency can be set to 1, 2 or 3 clocks. If a READ command is registered at clock edge n, and the latency is m clocks, the data will be available by clock edge n+m

Because of its excessive speeds DDR2 suffers from lax timings, however; the combination of an increased pin count (lower voltage) and Elipida's recent success with 90nm process bodes well for DDR2's future. Since testing DDR2 specifically 667MHz modules I've noted a propensity for latencies to increase running SPD (Serial Presence Detect) or Auto BIOS settings. While this is also true with DDR, at speed most modules maintain their CL (CAS Latency) rating. This may be due to the use of native 400MHz or 533MHz DDR2 IC's. Mushkin's Enhanced Memory PC2-5300 maintained their rated CL4-4-4-11 at speed running under SPD or Auto. I've provided thumbnails to a few screenshots 1st exemplifying performance at default speed/3.4GHz running under Auto.



Our second thumbnail represents the system at its highest stable overclock (tested running Prime95 and SETI consecutively) running 250FSB or 4260MHz at 1.151Vcore, VDimm = 1.95V. SiSoftware Sandra 2005 results at 6238MB/s are not as impressive as I would have liked, however; given the latency penalty the bandwidth isn't poor by any stretch of the imagination.



For the remainder of our tests all BIOS adjustments are on Auto with the exception of FSB which I had to decrease from 204MHz to 200MHz. This is an annoying attempt on behalf of motherboard maker to give their product an edge in performance. Given the fact the AA8 is perhaps second only to the Asus P5AD2 in performance as a i925X platform, this prima facie cheat is nonsensical.

So now onto our battery of benchmark ->