In the fast growing computing systems, memory selection for different applications is becoming more and more complicated. The rapid progress of the DDR technology and huge specifications differences among vendors make the general knowledge of memory architecture not enough. Engineers have to learn more details about the specific memory to choose the right one.
Understanding DDR Architecture and Generations
DDR memory is Double Data Rate memory. It transfers data on the rising and/or falling edges of the clock signal. Therefore, it is called Double Data Rate memory, or DDR memory as opposed to Single Data Rate (SDR) memory. Each generation of DDR memory improves in terms of speed, capacity and power efficiency.
DDR4 runs between 1.2V and 1.35V and up to 3200MT/s, while the newer DDR5 runs at a constant 1.1V but increases the MT/s to 3200 and up to 6400MT/s. As well as the increased MT/s per pin the DDR5 also includes on-die ECC for error correction as well as improved power management in order to reduce power consumption.
Memory Timing Parameters Matter More Than Headlines Suggest
Other parameters such as memory timing (CAS latency, tRCD, tRP etc.) play an important role in system performance. The headline numbers for memory often highlight the maximum speed that the memory can be driven at (in the case of DDR4 for example this is usually advertised as DDR4 3200MT/s) but the real numbers that define performance are found in the detail of the memory timings and the actual maximum speed at which the memory can be run (the so-called data rate). While lower CAS latency values are generally better, the lowest value does not always guarantee the best performance, it also needs to be combined with the highest data rate.
The two timing parameters tRCD and tRP are often are underestimated. As server applications access memory locations completely randomly, the memory has to access completely different memory locations (rows) every now and then. Here it is important to choose the optimal memory with the best timing parameters for tRCD and tRP.
Speed is not equal to Performance – Context is King.
Form Factors and Physical Considerations
There are many different Standard form factors for memory modules, often split between the dual-inline memory modules or DIMMs, and the small-outline dual-inline memory modules or SO-DIMMs. In terms of the pinout of these modules, the 288-pin Standard form factor for the DDR4 memory modules is also the 288-pin Standard form factor for the DDR5 memory modules.
Other issues to take into account are the physical height of the memory module. Depending on the size of the cooler on top of the CPU, it is possible that the memory modules would protrude too far from the motherboard board to be safely installed. It is therefore necessary to check the dimensions of the modules before purchase. The majority of modules are of normal height but there are also versions which are lower and which need to be specially secured in the slots with screws. Some modules are equipped with a heat spreader. These are no problem and are installed and removed in the normal way. There are also modules which need to be actively cooled. These are not suitable for normal use on a development system. In industrial use they are also installed in the normal way.
In industry we can recommend to use SK Hynix DDR memory modules. These modules are suitable for industrial applications, due to their proven reliability, stable operation and guaranteed specifications.
Error Correction and Reliability Features
Memory can also be equipped with Error-Correcting Code (ECC) for improved reliability. The nature of the ECC allows single-bit errors on every word of data to be corrected as well as any number of multi-bit errors on a 64b word. This makes ECC memory a requirement for many mission-critical commercial applications. Although processing overhead for ECC in commercial computing is about 5% in terms of required processing, for others the cost of re-running an experiment is far greater than additional processing to run the experiment with ECC-enabled memory. For example, for weather modeling, bioinformatics, etc. the need for ECC is high.
Within DIMM-based systems, there are two ways to increase signal integrity in multi-DIMM configurations: Registered memory modules (RDIMM) and Load Reduced DIMMs (LRDIMM). As previously mentioned, RDIMMs add a register on the memory controller interfaces with in order to reduce the electrical loading on the module’s pins. The LRDIMMs, on the other hand, add a buffer between the address/command and data signal pins and the DRAM module. This configuration primarily benefits the data signals in multi-DIMM configurations, whereas the register on RDIMMs can also improve signal integrity on the address/command signals as well.
Compatibility Challenges That QVL Lists Don’t Solve
But in reality memory compatibility goes far beyond just checking the generation of memory (e.g. DDR4) as well as the physical form factor (e.g. DIMM vs. SODIMM). Even high quality modules do not automatically have to be on a qualified vendors list (QVL) of the main
Remember, the purpose of the QVL is to give the customer a “head start” to start selecting a proper memory for his system. This list of qualified memory modules for a given motherboard will include tested combinations of various vendors and speed grades of memory. There will also be many quality memory vendors not included on the list. It is also very important to note that even quality memory can fail due to incompatibilities with a system. The only way to verify the proper operation of a given memory module is through testing in your system.
JEDEC, the standardization organization for memory, established various parameters for the different types of memory and for the different data rates. Furthermore, there are non-standard memory profiles (in addition to the standard JEDEC memory timings) that memory modules support in order to achieve higher data rates and/or even tighter timing parameters. If a mainboard and its BIOS is then also supposed to support such enhanced memory profiles, for example in the form of XMP profiles (for mainboards with Intel CPUs) or of DOCP/EXPO profiles (for mainboards with AMD CPUs), then these can be easily enabled via the BIOS settings of the respective system.
Application-specific Selection Criteria
An example would be high frequency trading platforms. Here the lowest possible latency would be required to be able to process large amounts of information in real time. Other applications such as video editing would profit from high amounts of memory to run large numbers of projects at the same time and do them as fast as possible. For gaming it is typically the case that the fastest memory with the lowest possible latency leads to the best performance. As already mentioned, however, at a certain point additional speed does not bring any noticeable improvements anymore.
In order to assess the Power Consumption of memory it is necessary to also take into account the planned area of application. Could the system be used in a mobile environment or in a very densely packed server environment? If so the lower operating voltage of DDR5 could result in significant power savings, even though in small deployments the higher cost to transition to DDR5 memory may not be justified. The Operating Temperature of memory is also very important for industrial applications. Standard commercial memory is typically specified to operate up to 85°C but there are also industrial-grade versions of memory that can operate over extended temperatures.
Planning for Memory Technology Transitions
Even if a system designer wants to keep his or her memory-based applications up to date with newer technologies that should offer better performance and features as well as lower cost per capacity, that designer will want to wait until the newer technology goes into full volume so that it can offer better performance as well as lower cost per capacity for the newer technology than the currently-fully-adopted technology is able to offer.
Industrial customers plan to use memory over a long period of time and therefore must have an eye on the memory eco-system. Established memory technologies are available at the best price performance in their middle of life. For instance CPU’s and chipsets for servers and storage already support DDR5, while the prices for DDR4 modules are rising because of the growing volume of DDR5 modules. That is why established technologies like the widely used DDR4 are the cost-effective alternative for applications that do not require the highest performance. Typically the window of best price performance lasts for 18 up to 24 months until the availability of the technology starts to be constrained.
