DDR2 - SH/SC Wiki

What is DDR2 memory

by Dr. Fred

Double Data Rate Synchronous Dynamic Random Access Memory 2 (or DDR2) is the next evolutionary step up from standard DDR. It brings several enhancements to the table, along with greater bandwidth and the ability to hit much higher speeds. The fastest DDR1 overclocking RAM currently roofs out at around 600MHz (~PC4800), whereas DDR2 is hitting over 800MHz (~PC2-6400) and has the potential to go much higher.

What makes DDR2 different? Well, first of all, core clock speed is halved in favor of 4-bit data prefetch, compared to 2-bit for DDR1. This means that for, say, DDR2-400, the core clock speed is 100MHz and the data transfer speed is (100,000,000Hz x 4 bits ) x 64 bits / 8 = 3,200 million bytes per second. Data is still transferred twice per clock cycle however, as the I/O buffers send data to the memory bus at double the core clock. Putting it all together, the memory controller runs at 200MHz, receives 2 bits of data every half clock cycle - same as DDR1 -, but the clock frequency of the memory chips themselves is halved. This not only enables higher speeds, since less strain is put on the memory chips, but also plays a part in cutting down power consumption. Partly because of this, DDR2 DIMM slots have higher pin density than DDR1 ones (240-pin compared to 184-pin), although the slot width remains the same.

DDR2 also incorporates three technologies aimed at improving signal integrity, lowering power consumption and thus allowing higher clock speeds: ODT (on die termination), ODC (off-chip driver calibration) and FBGA (fine ball grid array).

ODT simply terminates signal reflections directly within the memory chip, as opposed to DDR1 where they must be driven down to a resistor on the motherboard, causing extra disturbance in the process. This helps reduce signal noise, and also happens to cut costs for motherboard manufacturers.

ODC is a little more technical, but essentially it allows the BIOS to adjust buffer impedance in order to keep bus and core strobe signals synchronous. In English, that means it basically attempts to keep different clocks synchronous with each other. This improves signal consistency, which also helps compatibility between different memory modules and memory controllers.

Finally, FBGA is simply a different means of connecting the memory chips to the module PCB. It offers much better electrical and thermal characteristics than the traditional TSOP (Thin Small Outline Package) format, as well as the possibility to stack multiple dies in the same memory chip to increase capacity. FBGA has in fact been used by some manufacturers for DDR1 modules, but is now a requirement for the DDR2 spec.

Combined with halved internal clock speeds, these technologies allow the nominal DIMM voltage to be reduced from 2.5v for DDR1 to an impressive 1.8v for DDR2. This greatly diminishes power consumption despite the much higher data speeds, and cuts down on heat dissipation as a bonus.

DDR2 also uses Additive Latency, a method of issuing commands designed to prevent conflicts or collisions between queued instructions and improve overall bandwidth. Unfortunately this doesn't really help the fact that, because of its low internal clockspeeds, DDR2 has much higher latencies than DDR1. To compare, low-latency overclocker DDR1 modules can go down to 2-2-2-5 timings, whereas DDR2 will not go below 3-3-3-8. In practice this means that, at the same rated speed (say DDR400 and DDR2-400), the DDR2 memory will be slightly slower than DDR1.

Then there's the fact that, despite the DDR2 support in 915/925 motherboards, Intel only makes two processors that actually use the extra bandwidth DDR2 offers: the [at the time of writing] $999 3.46GHz and 3.73GHz "Extreme Edition" Pentium 4s. Current even the fastest Prescott chips use Intel's quad-pumped 800MHz FSB, meaning they're only capable of transferring 3200 million bytes of bandwidth per channel per second, which PC3200 DDR1 already fulfills.

Now in reality, the memory controller does benefit from running faster DDR2 RAM (i.e. PC2-533 or PC2-667) asynchronously with the 800MHz FSB, but the translation to real-world gains is very small. Even the 3.46GHz Extreme Edition chip, which uses a 1066MHz FSB (and can thus fully utilizes DDR2-533 RAM), is barely any faster than its 3.4GHz/800MHz FSB predecessor. It is also beaten by Intel's own $800 3.8GHz Prescott chip, which gets by just fine with good old DDR400.

Considering the higher pricing compared to DDR1, one really has to wonder whether DDR2 is even worth it at all. For overclockers, it obviously allows extremely high speeds with low heat dissipation, and as such can be preferable over high-end DDR1. For the consumer, I would have to say "not yet"; at least not until some 1333/1600MHz FSB chips come out, or DDR2 prices end up evening out with DDR1.

DDR2 will be adopted by AMD with the new Socket M2 A64s and A64 X2s some time in the first half of 2006, by which time prices should have lowered and the new dual-core chips will provide a real justification for the extra memory bandwidth. Intel's next generation mobile Pentium M chips also use DDR2, which helps improve battery life compared to DDR1. Looking further ahead, Intel plans to implement DDR3 by 2006-2007, so DDR2's success could be relatively short-lived. Considering DDR2 was originally scheduled to succeed PC2700 two years ago, however, DDR3's timeliness and availability remains to be seen.