Why switch to DDR4?

^{(a source)}
Not so long ago, a standard for DDR4 SDRAM, a new generation RAM, was published. The standard describes in detail the memory device, but what, in fact, differs from the memory of the previous generation, or what advantages does the transition to DDR4 say or very little or not said at all (probably in the hope of readers' imagination :-)) .
In this post I will try to explain the main differences between DDR3 and DDR4, and what advantages the new memory standard has for the end users.

Increase volume and performance

One subtle, but at the same time important difference is how the memory chips are organized.
The 8Gb x4 DDR4 chip usually consists of 4 groups of banks, 4 banks in each group. Each bank of this chip contains 131.072 (2 ¹⁷ ) rows (rows), 512 bytes each. For comparison, the 8Gb x4 DDR3 chip contains 8 independent banks, 65.536 (2 ¹⁶ ) lines per bank, 2048 bytes in each line. With an equal volume, the DDR4 chip has twice the number of banks and is much shorter than the memory line. This means that the new memory can switch between memory banks much faster than DDR3 did. In particular, for 8Gb x4 DDR4 chips declared as 1600 MT / s compatible, the tFAW (Four-bank Activation Window) indicator is 20ns, which is half the size of DDR3 (40ns). This means that DDR4 memory chips can open arbitrary lines in different banks twice as fast as DDR3.

A comparison of DDR3 and DDR4 shows that the largest DDR3 module that can theoretically be constructed will be 128GB (using QDP (quad die package) and packing four chips in one case) and 8Gb crystals (Fig. 1). For DDR4, using 16Gb crystals, and eight-layer packing of crystals in the chip, it is theoretically possible to create a memory module of up to 512GB. The number of contacts on DDR4 modules increased to 284 to address this amount of memory. Each DDR4 memory chip can be a stack of 2, 4, or 8 DRAM crystals. The 8-layer stack is described in the specification supplements and is likely to require the use of TSV (through silicon via) for its practical implementation.
In general, all these changes are aimed at creating larger memory modules and increasing performance.
')

Figure 1. Comparison of DDR3 and DDR4 modules ^(source)

Improved energy efficiency

Another important part of the DDR4 specification is energy efficiency improvement compared to DDR3. In addition to reducing the voltage to I / O from 1.35V to 1.2V, the new standard also specifies the use of a higher voltage level inside the chips (DRAM word line 2.5V), which provides fast access in active mode and low leakage current in the passive mode.

The electrical implementation of the I / O interface has also changed. The new interface is called pseudo-open drain (POD, “pseudo-open drain”) and its main difference is that current does not flow in the circuit when a high voltage level is set on the line. Electrical interfaces DDR3 and DDR4 are shown in Figure 2.


Figure 2. Electrical data I / O interfaces for DDR3 and DDR4. ^{(a source)}

Reducing the voltage on the I / O, changing the electrical interface and reducing the length of lines in the memory banks lead to a significant reduction in power consumption compared to DDR3. Preliminary estimates say 30% winnings. Although, of course, it depends on the nature of memory accesses, technical process and many other factors. Such a gain can be used to increase the clock frequency and, accordingly, the speed of work, or to save a little energy with the same performance.

Reliability

Also, many significant changes relate to the reliability ( RAS ) of DDR4. For example, the specification says about the detection and correction of memory chips errors associated with the control of the parity of commands and addresses.

Another example is that DDR4 chips have a connection test mode. This mode allows the memory controller to check electrical connections (and find line breaks) much faster than before, and without the use of initialization sequences.
Also, DDR4 can be configured to discard commands containing parity errors. In DDR3, such teams were skipped and reached the memory chips, repeatedly complicating the recovery from failures.
And as an example of one of the optional “features” that the specification contains, you can give a check of checksums for the data stored in the memory.

All these and other features are designed to ensure the growth of working frequencies and memory volumes (associated with an increase in the number of errors in the work), while ensuring stable operation.

With higher performance and energy efficiency, DDR4 memory already in 2014 should easily take its place in multi-core server and desktop systems. And then, at the expense of a lower price per unit of volume in addition to other benefits, DDR4 should get to other devices.