A guide to emerging data memory and storage technologies
Technology breakthroughs are changing the way we live and work and with that change comes a change in datasets. The amount of data we produce at scale is significantly larger than decades ago and enormous amounts of data require rapid development of memory storage technologies and hardware upgrades. In-memory computing, new memory technology, and secure infrastructure are some of the biggest considerations.
The delays in moving data between memory/storage and compute in conventional computer architectures makes it difficult to deal with big data application such as AI and the IoT. This is resulting in the move to new computing architectures that bring processing closer to data.
Among the approaches being pursued are the use of in-memory computing platforms for cloud and data center memory and storage management, which minimizes the movement of data for processing.
Efforts to move processing closer to data is also leading to new hardware and interface developments, such as various types of computational storage that brings specialized accelerator processors closer to the data they work on for better data management, but also for an increasing number of more conventional processing applications.
For example, the new Compute Express Link Interconnect (CXL) is a new high-speed interconnect designed to accelerate next-generation data center performance. This runs on the Peripheral Component Interconnect Express, or PCIe, physical layer, which will make PCIe gen 5.0 and 6.0 able to handle cache coherency and management of various processor accelerators such as graphics processing units (GPUs) and Field Programmable Gate Array (FPGA) based specialized computing devices.
CXL is intended to be a high-speed open standard interconnect that enables an accelerator ecosystem geared towards high performance, heterogeneous computing. CXL will likely also work with other network interfaces that aim to bring in-memory (or in-storage) processing to larger storage system and data center applications. The image below shows CXL between the PCIe physical layer and the data center network.
True in-memory processing with an intimate connection between memory and processing are still some years away and may use spin-based logic in conjunction with magnetoresistive random access memory (MRAM) memory cells.
New non-volatile memory technologies are poised to replace volatile memories and enable new applications both in the data center and bring new materials into semiconductor wafer production.
Resistive random-access memory (RRAM) and premium computer support (PCM) technologies are also being used to create neuromorphic computers that mimic the performance of the synapses in the human brain. These could be useful for machine leading (ML) training and may use less power for such training than graphics processing units, for example.
On the application side, inference engines running on embedded devices could benefit from the use of non-volatile memory to store weighting values. Additionally, because of scaling issues with NOR flash, MRAM could replace embedded NOR in many applications.
At least one company has replaced lower performance static random-access memory (SRAM) with MRAM because the SRAM memory cells are much bigger than the MRAM cells and allowed more memory for a given die size.
There are also efforts to use crossbar-based RRAM memory in AI applications. All of the major silicon foundries are offering MRAM or RRAM memory options in their embedded system-on-chip (SoC) products.
In the future neuromorphic memories could be an important element in AI model training and non-volatile memories will replace many volatile memories such as dynamic random-access memory (DRAM) and SRAM, resulting in lower power consumption that will be a great benefit for battery powered devices. An emerging MRAM technology, called spin orbit torque (SOT) MRAM, would have latency and write/read speed that rival those of SRAM, perhaps enabling the use of non-volatile memory in most memory applications.
New computer architectures will create new intrusion vulnerabilities. For instance, moving from volatile to non-volatile memories means that the data stays in the memory even if the power is off. This makes these new memories more vulnerable to data extraction than was the case for volatile memory systems. Strong encryption and special data wipe capabilities will be needed for systems that storage data in non-volatile memory to avoid such attacks.
As computers become more sophisticated modern encryption technology could be broken. For instance, successful quantum computer technology could break many current strong encryption algorithms. For this reason, new encryption technologies are needed.
Leading IT service companies have already begun to develop cryptographic algorithms that are resistant to potential security concerns posed by quantum computers. Such strong quantum computing proof encryption technologies could become a standard in storage and memory technology in the next few years.