There are some statistics that just makes your brain go numb. 3.2 billion transfers per second is one of them. It indicates the maximum speed DDR4 RAM modules transfer data to CPUs. This means that on an average, they would be twice as fast as the fastest DDR3 memory. The first modules for PCs will have a transfer rate of only 2133MHz, but even this is already double of what DDR3 modules in the latest PCs today offer – only 1066MHz. Even the maximum permissible storage capacity per module will increase: sizes starting from 4GB will become standard. What’s more, there will be less power consumption thanks to a lower operating voltage.
To make this performance leap possible, the Joint Electron Device Engineering Council (JEDEC) has also made some outward changes in the module: a fourth generation RAM module has more pins that connect it to the board to fulfil the high speed requirements. Each DIMM has a direct line to the CPU controller to prevent delays and/or bottlenecks in the data transfer from CPU to RAM. This point-to-point concept will replace the multidrop bus, on which the individual modules are currently arranged.
The next RAM generation offers double the speed and can store much more data than ever before. A look at the new module shows how this is possible
The DDR4 RAM thus doesn’t fit into the current boards and requires new CPU controllers. We can expect both to hit the stores at the end of the year when Intel launches its high-performance CPU Haswell-E that uses the DDR4 RAM. However, the new RAM for inexpensive PCs will not be launched before next year when Intel introduces the Haswell successor, Broadwell.
DDR4’s bank secret
Majority of innovations in the DDR4 are related to its internal structure. A DIMM is made up of memory chips that are controlled with a controller. Banks are the central units inside a chip. The performance is finally determined by how data is read from or written in a bank (refer to images on the left). During every reading or writing process, the bank always opens a word line and all connected transistors lose their charge – each charge corresponds to a bit. These bits have to rewrite the signal amplifier of every bit line. In DDR4, the charges (bits) of the banks are extracted faster when they are full because a word line holds fewer bytes than those of a DDR3 module. The increase in speed can be measured: the time specification tFAW (time Four Bank Activation Window) describes how quick word lines can be opened in four banks. For a DDR4 module, the tFAW is 20 nano seconds (ns), while for a same-sized DDR3 RAM it is 40 ns. The length of word lines has been reduced to a quarter in DDR4 but their number has been doubled. Therefore a bank stores only half as much data as DDR3. On the other hand, a DDR4 chip has twice as many banks, which are also integrated into groups: each bank group contains four banks. The organisation in bank groups allows rapid transmission of data from one bank to the temporary storage (prefetcher) and finally to the data bus. This transmission gets more and more complicated with every new RAM generation: the clock frequency, at which the signal amplifiers of bit lines work, is maximum 266MHz since the DDR2 days. To achieve higher data rates, manufacturers increase the parallelisation with each RAM generation: one DDR2 memory chip collects data from four banks and DDR3 captures data from eight banks in the prefetcher, which transmits data with a higher clock frequency. DDR4 would have bagged more points with as its prefetcher is two times larger and is therefore a wider data bus right up to the CPU. The JEDEC wanted to avoid this and therefore did not increase the size of the chip prefetcher. Instead, it has provided an upstream intermediate storage along with the prefetcher of the bank group. This considerably increases the clock frequency while the width of the data bus remains the same.
Less consumption, more stability
DDR4 is also equipped with new features that reduces its power requirement and increases its reliability. The first thing that contributes to saving energy is the shorter word line, because it needs less power to get activated. This enables a reduction in the operating voltage from 1.5 (DDR3) to 1.2 volt (DDR4).
In addition to that, the memory chip has a separate power supply, which it can increase the operating voltage to 2.5 volt if required. At the same time, a DDR4 module can dynamically adjust its refresh rate. Unlike Flash memories, RAM cells cannot store their content permanently. The content needs to be refreshed at certain intervals depending on the operating temperature – at the latest after 64 milliseconds. These intervals are defined when it comes to DDR3, while a DDR4 module refreshes the content depending on the temperature and can reduce the refresh rate to save energy. This feature alone reduces the energy requirement of DDR4 by 20 percent when compared to DDR3. In total, the energy saving is between 30 and 40 percent, as estimated by the JEDEC. Another change is the integrated error detection and correction, which is much faster compared to DDR3. There is another future-oriented feature, which is stipulated in the standard and which no one can implement today: with the help of chip stacking, manufacturers can stack RAM cells in up to eight layers on top of each other which dramatically increases the memory thickness and enables modules of up to 512GB – one such giant module would be larger than most of the latest Flash-based hard disks.