However, things do change - GeIL used Computex Taipei 2006 as a place for releasing its fastest memory to date, which just happens to be the fastest specc'ed DDR-II memory modules out there. The retail package contains two memory modules with 1GB per module, and the company can thank the memory performance to their newly-developed Dual Channel Making Machine, a memory sorter which enables hand-picking sort of all 32 chips that are spreaded on two DIMMs. Once the sorter does the job, the ideal configured chips are being glued to the PCB and then retested again. End result is "dual-channel" memory from the memory chip onwards.
Too good to be true?
The memory comes in a compact retail package, which only shows two aluminium heatspreaders painted in bright orange. Folks from Netherlands will probably just love the Oranje colour. However, this memory isn't for those that want fancy-smancy designed modules, these babies are bred for one thing and one thing only - speed. These modules could be compared at first with a Lambo LM002 - brutal, with no finesse. The memory is officially named DDR2 Ultra Low Latency Multi-Spec Dual Channel Kit. Very short name, as expected by the Taiwanese companies - we still remember mind-boggling name convention from Gainward. The Multi-Spec part of the name is perhaps the most intriguing one, since this memory isn't a follower of Corsair-Nvidia EPP camp. Rather, the company has memory specc'ed at PC2-8500 if you run Intel and astonishing high PC2-9600 if you're running AMD platform. Dr. Jekyll gives you a bog standard 1.06 GHz clock with CL5, 5-5-15 latencies. But if you plug the memory into the DIMMs next to Socket AM2, you're starting to toy with Mr. Hyde.
The declared spec is 1.2 GHz (600 MHz DDR) at CAS4 4-5-10 latencies. At 800 MHz, the memory is specc'ed to work at CAS4, 4-4-12-2T - but that was far from the truth. We got the modules running with zero problems at CAS3, 3-3-9-1T on both Intel and AMD platforms. This was achieved on more than one motherboard and we managed to get confirmation that this isn't the isolated case with our sample.
In order to test the memory right, we are used two different configurations: AMD and Intel are both represented, and this is the default configuration for testing memory modules in future. With Intel Conroe platform, bear in mind that the memory scores have less impact on system performance since Chipzilla used all of its magic to stop the CPU from reaching to the ancient 64-bit GTL+ bus, which connects the CPU to the NorthBridge chip which hides the 128-bit, dual-channel memory controller.
Testing configuration #1 - AMD Athlon 64 FX-62 CPU, Foxconn nForce 590 based motherboard, Nvidia GeForce 7950GX2 graphics adapter, 250 Gig Seagate and Hiper 580W PSU. DVD device was Samsung SATA DVD toaster. Monitor connected to the graphics card was four year old Dell's 19" LCD baby.
Testing configuration #2 - Intel EE 955 CPU and Core 2 Duo X6800 switching places on a D975XBX motherboard, Nvidia GeForce 7800GTX graphics card connecting to 19" Samsung 959NF CRT. All is powered by a Tagan EasyCon 580W PSU. The storage subsystem is consisted out of 250 Gig Seagate and Plextor PX-760A DVD toaster.
Both machines were running clean installs of Microsoft Windows XP with SP2 and all updates Microsoft can think off, plus the latest ForceWare drivers from Graphzilla. Benchmark suite was consisted of SiSoft Sandra 2007, Everest Ultimate Edition 3.0 and RightMark Memory Analyzer 3.67. We had great expectations of this aggressively specc'ed memory, and it did deliver. Since GeIL specifies that the memory is best utilized with AMD AM2 platform, we used the AMD platform first. Getting the memory to run at PC2-9600 was a breeze. Enter BIOS, getting the CPU set at 10x300 and leaving the default memory divider was pretty much enough. Of course, memory was manually set at 2.4V, which is mandatory if you wish to avoid the beep beep sound coming from the speaker.
Our sample worked at 600 MHz DDR (1.2GHz) with CAS4, 4-4-15-2T latencies. Sandra 2007 reports bandwidth in excess of 11 GB/s - Integer was 11.45 GB/s, while Float memory performance hovered around 11.44 GB/s.
When testing the memory with Everest Ultimate Edition, the read speed was set around 10.2 GB/s, which is highest score in this app - even dual Socket Optys do not score this high. Test which I consider to be more important that read one is the write speed. Many of memory manufacturers were simply dissing the write bandwidth if their new-super-duper memory didn't achieve a good score in this area - this is also test where you can see the difference between various latency settings. From this MultiSpec kit, you can expect a cool 8.57 GB/s of write goodness. GeIL's high clock also has lowest latencies I have ever witnessed in DDR-II world - 37.5ns is pretty much on same levels as DDR-I memory on older DDR-I powered Athlon 64 X2 4800+. If you cannot run the memory at 1.2 GHz because the CPU or the motherboard does not allow it, you can expect a stable frequency of 800 MHz with CAS3, 3-3-9 latencies. On ASUS motherboard, we had zero problems with getting the memory to work with Command Rate set at 1T. Foxconn motherboard, however - does not want to work with 1T, so we had to settle with 2T. Working with DDR-II memory at DDR-I latencies at double clock proves a point that AMD was right with its decision to skip the first two generations of DDR-II.
In La Intella world, the latency of CAS5, 5-5-15-2T at 1.06 GHz was really too conservative. These modules have no issues working at CAS4, 4-4-12-2T at 1.06 GHz or CAS3, 3-3-9-1T at 800 MHz. This was achieved on Intel's Bad Axe motherboard, so no ultra-super-duper overclocking motherboard with special Oscar-winning BIOS. However, Bad Axe is far from being a bad motherboard - rock stable operation of Conroe at 3.53 is easily attainable if you have the right cooling. The scores this memory was achieving were around 300-400 Mbytes per second faster than comparing Corsair XMS22048-8500C5, but the difference in access time is close to 10 nanoseconds (51.47ns compared to 59.74ns). The bandwidth numbers of Intel Conroe are around 4.5-5.5 GB/s, so don't expect world breaking scores here.
In case of this memory, the Dr. Jekyll part was completely suppressed by Mr. Hyde. If your motherboard is able to supply 2.4 Volts, you can start toying around with latencies that go well bellow the specs that are declared on the memory, which is something we haven't seen since Corsair's famous XL Pro modules that worked on CL1.5, 2-1-4 at 3.1V, but only when a special link between Earth, Moon and Mars were established. GeIL caught us by surprise, since. The only negative point is the fact that modules require extremely high memory voltage. We're talking about minimum 2.3, 2.4 or 2.5 Volts of working memory, which is something that we saw on a very small number of motherboards. However, if you're buying this memory, chances are you aren't looking for anything but the best - DFI LANParty nF590SLI, universal abit AN9 32X Fatal1ty, ASUS M2N32-DeLuxe or something similar. Crew from GeIL told us to expect fireworks with Intel platform after DAAMIT RD600 chipset comes into frame, later this month.
Even now, two months after the launch, there aren't any modules from competing companies that can work with settings like these. We're expecting an answer from Corsair and other players, but for now, GeIL rules the roost. If you want the fastest memory out there, don't look any further than two Oranje sticks. ?
Sign up for INQbot – a weekly roundup of the best from the INQ