Kingston Hyper-X3200: DDR400 With a Healthy Identity Crisis

Kingston has been providing the PC market with reliable, cost effective memory for some time. Their Hyper-X series has been a favorite of the Overclocking and Enthusiast communities, as well the average PC-user desiring a higher performance part. Since the introduction of the dual memory controller, especially Intel's Canterwood/Springdale chipsets, Hyper-X branded memory has been coveted among the overclocking genre.

Scanning through Enthusiast forums reveals Hyper-X to be a staple among the few brands able to meet the demands of this discriminating segment. The overclocking community in particular, is much more critical of their hardware then the perhaps any other. By definition overclockers seek only those products capable of far exceeding the manufacturer's specifications. Their devotion is not limited to any one brand, but to the best performer. In the memory market, this may change from one moment to the next.

When the I875 chipset was introduced, Kingston Hyper-X was manufactured with Winbond BH-5 IC's (Integrated Circuits). Since Winbond stopped production of these integrated circuits, tales of their existence have attained a somewhat Homeric mythological status. Samsung Semiconductor has given us DDR which comes very close to surpassing the performance of BH-5, running at CAS2 2-2-5 under several manufacturers' heat spreaders. The difference between Samsung, and Winbond DDR-IC's, is the latter can soak up voltages as high as 3.45V, and I've heard rumors this has increased. This allows the memory to run at very high frequencies, while maintaining very low latencies.

In the memory tested today, Kingston hasn't necessarily endeavored to recreate the performance found in its former BH-5 supplied Hyper-X. In fact, given the Overclocking community comprises a mere 5% to 11% of total memory sales, producing such a Low Latency memory in 2048MB kits, would also be mythological, but more along the “Homeric” Simpson, at least financially.

The memory tested today, while exhibiting attributes innate to the Hyper-X series, also maintains the cost consciousness which makes Kingston an industry leader. Unfortunately for many Enthusiasts, 2048MB-kits are most often narrowly associated with Server, or Large 3D applications. What I hope to communicate today, beyond this memory's performance potential, is that the evolution of the modern PC is perfectly complimented by a kit of this size.

The Hyper-X3200 modules run at a latency of CAS 2.5 3-3-6. To manufacture extremely low latency modules of this size would have been far too costly. A cost Kingston is unwilling to pass on to the consumer. Hyper-X modules most definitely represent high performance, regardless of kit size. As you will see in this review, this particular kit performs as if it were a whole other memory, which was the inspiration for my title.

I've chosen the Asus P4C800E-Deluxe motherboard as our test bed, largely because it's the only 875-chipset based motherboard I own (please send boards Asus, Abit) and its MCH (Memory Controller Hub) is capable of reaching very high front side bus speeds. The Canterwood chipset is an able overclocker and will make an excellent catalyst for our processor, the largely misunderstood Socket-478 Prescott used in this test.




Our test system today, will include the Asus P4C800E-Deluxe (Beta BIOS ver.1017), Socket-478 Prescott 3.0E (SL79L Philippine), and a Sapphire X800Pro. The system is water-cooled, via the Cool-Cases CC-Magic (CPU-kuhler), and Danger Den Maze-4 watercooling the R420 GPU. The CPU runs completely stable up to 240FSB (1:1) under Default Vcore (1.365V ~ 1.475V via D-VID), and has been overclocked to 4010MHz (267FSB), although the system has not been completely 3DMark stable above 3.9GHz (yet). I chose an 875-chipset motherboard for its versatility in overclocking, specifically its high FSB limit. Our 2048MB pair of Kingston Hyper-X sports the usual blue anodized aluminum heat spreaders, which give Hyper-X its attractive trademark look. The lightweight aluminum seems much more capable of dissipating heat then the atypical "Gold" heat spreaders found on other brands.




Memory kits of this size are not strictly for Servers, CAD/CAM, or other such graphic intense applications. Your PC works on a memory hierarchy, after the L1, L2 caches are exhausted, and your physical RAM depleted, it's the system's pagefile allocated to your hard drive by the operating system which handles the excess. The pagefile, Swap-File or Virtual Memory as it's known, acts almost as an external cache, although that's a loose definition. When nVidia introduced its nForce chipset, one of its most ingenious attributes was given the moniker DASP (Dynamic Adaptive Speculative PreProcessor). This feature which resided in the North Bridge, identified and anticipated instructions or data which was most often shared between CPU and main memory. The information included some graphics related instructions, and DASP was also described as an L3 cache. I've been experimenting with several pagefile settings, finally disabling the pagefile all together. The performance is noticeable as the PC seems more responsive transitioning between programs. Disabling the pagefile does not necessarily enhance performance in so far improving benchmark scores. Onto the details of our test system.

Socket-478 Prescott 3.0E SL79L

Asus P4C800E-Deluxe (BIOS Beta Ver.1017)

Kingston KHX3200K2/2G

Sapphire X800Pro (Catalyst 4.7)

Maxtor Diamond Max Plus 9 (SATA150 120GB)

PCPower&Cooling TurboCool 510 Deluxe

TTGI USA TT-201T3

WindowsXP SP1

The CPU-Z screenshot below indicates memory's behavior under the SPD command at default (200FSB 1:1) speed.




As Kingston specified we see above a CAS Latency of 2.5-3-3-6, at 200Mhz (200FSB 1:1) or DDR400. In the our next CPU-Z screenshot below the CAS-latencies are raised to CL3-4-4-8 under the SPD command, at 250MHz (250FSB 1:1) or DDR500.




The aspect ratio will remain at 1:1 for all our tests, and the DDR-voltage (VDIMM) will remain at 2.75V, increasing to 2.85V only at 250FSB (DDR500). For our first set of benchmarks I employed SiSoftware's Sandra Professional memory module, in both buffered and un-buffered modes. Since testing on Intel's Canterwood/Springdale platforms, I've found there's a significant improvement in performance, whenever the SPD (Serial Presence Detect) command is enabled over Manual settings. Fortunately the average PC-user will have a minimum of BIOS adjustments to concern himself/herself with. As indicated above by CPU-Z, our KHX3200K2/2G ran at CL2.5-3-3-6 at the default speed of 200MHz, or DDR400. Anything above this, and SPD automatically raised the CAS-latency to CL3-4-4-8. I did try running the memory at a tighter CL2 by increasing DDR-voltage, however; I wasn't able to reduce latencies. I was, however; very impressed the memory ran as fast as 250MHz or DDR500 attaining PC4000 performance given the kit size.

As this particular memory doesn't fall into the recent genre of “Xtreme Series” kits designed to run from DDR400 to DDR500, such as Corsair's Twin-X1024 3200XLPRO reviewed here.



The bandwidth is certainly impressive given the memory's default latency setting. There are a few variables which will inevitably affect bandwidth during these tests, including the prima facie enhancements, such as running PAT in Turbo mode, therefore I've left this setting on Standard. The memory performed completely stable through 250FSB, at least through Sandra's memory benchmarks. Our next series of benchmarks feature Aida32 system benchmark. Specifically it's READ and WRITE measurements. Aida32 usually produces much higher results then SiSoftware's Sandra Memory module.



Another benchmark utility, which memory will effect, is PiFast. For our purposes I chose Hexus Pifast Challenge as it is most likely the largest database of its kind for comparative results. Those whom participate, usually provide a detailed description of their system, I find the information invaluable. I included screenshots of results from default speed, and the highest stable overclocked speed, at 200FSB, and 250FSB respectively.





Certainly there's a significant difference in the Calculation of Pi at higher frequencies, and CPU speed is as important to this benchmark as memory commands. To run this benchmark yourself simply follow this link Hexus PiFast Challenge. Our next graph exemplifies FutureMark 3DMark03, and 3DMark2001SE benchmark results from each representing 200FSB to 250FSB system speeds. I ran the memory ran under SPD command, with PAT in Standard mode. VDIMM began at 2.65V from 400MHz (DDR) to 2.85V at 500MHz (DDR). CAS-latencies under the SPD command ran from 2.5-3-3-6 at DDR400 to 3-4-4-8 at DDR500.

Kingston's Hyper-X3200 in 2048MB-kit certainly complimented the performance in the system tested. I experimented with a number of Pagefile sizes, and found disabling the Pagefile to be the best option for this particular system. I have read there may be programs which expect to see a Pagefile, and if it's not present will not work properly. The cache is a layer between the kernel memory management code and the disk I/O code. When the kernel swaps pages out of a task, they do not get written immediately to disk, but rather are added to the cache. The kernel then writes the cache pages out to disk as necessary in order to create free memory. (Swaping and the Page Cache) Since disabling the Pagefile over a week ago, I haven't experienced any system anomalies. Disabling the pagefile effectively eliminates hard drive access as a memory related function. Disabling your Pagefile doesn't necessarily improve performance, there is, however, less time transitioning between programs, and simultaneous process are unaffected. In so far as the “need' for a kit of this size, considering the concurrent evolution of operating systems, and CPU's physical memory requirements will most likely grow. From the Windows 98 operating system, to Windows XP, the amount of physical memory allocated to the OS only has doubled.

There's a propensity among chipmakers and software developers in which server technology has permeated the designs of Desktop Computers. Many would claim the minimum memory requirement for a Windows XP based system to be adequate at 512MB, while in truth the operating system itself can consume 256MB. With the advent of Hyper Threading, and other technologies which seek to emulate a dual processor environment, physical RAM is taxed even further. For the modern PC, memory is still the fundamental principle on which its CPU is based. Dissecting today's processors such as the Athlon FX, reveals the implementation of Server technology at its core. Somewhat indicative of this design philosophy are the rapidly growing cache sizes on today's chips.

For the epitome of cache enhanced processing power one need look no further then the Intel Xeon MP. While the Xeon MP may use just a 512KB L2, it uses up to a 4MB L3, and this is not the usual throttled down L3 we've seen in the past. If we take into account most of today's high-performance processor's such as AMD's Athlon FX-53 now dominating the performance charts, it's apparent these designs are extrapolations of Server technology.


Unfortunately there's a propensity among many PC-users, in which their forever rationalizing inadequate amounts of physical RAM based on cost. The average person in the market for a PC, will usually base their decision on hardware compliments found on the larger PC-builders, machines. Many fail to realize, even when these companies build their flagship model, their cutting cost wherever possible. Basing your custom PC on the purely economic model of OEM PC-builders, would be analogous to the Sports Car Enthusiast, inserting an economy engine on a Porsche chassis. Not only is memory critical to system performance, but it's one of the least expensive ways to improve performance.

Perhaps there's no stronger evidence for what constitutes the ideal amount of physical RAM then to study the specifications of recent chipset's, and the motherboards their built upon. Examining the following Chipset Parameters for early Pentium-4, and Athlon-XP processors, show most of these DDR motherboards feature chipset's with MCH's (Memory Controller Hub) capable of accommodating up to 4GBs of RAM. If most of us were to use just 50% of the maximum allowable RAM for our current motherboards, the majority of us would be running 2048MB kits. About 100% more memory, then the amount found in the majority of today's higher performance PC's.

In the kit tested today, Kingston has provided us with another quality product far exceeding its default specifications. I do believe we've inadvertently stumbled across an overclocking sleeper, and in an area where I least expected to find it, a 2048MB kit.

Considering the cost (kits of this size are now under $700) we no longer have to sacrifice performance for price. Kingston's KHX3200K2/2G kit, would compliment just about any system's performance. Albeit gaming, overclocking, business applications, and/or server related tasks, this memory is rock solid, and highly overclockable. The kit can be found for as low as $659 at Newegg, (USA) and it comes with Kingston's Lifetime warranty.

I'd like to thank Heather at Kingston. Keith Suppe aka Liquid3D


Questions/Comments:forum thread