MEMORY Non- Technology Continue to Scale Up The emergence of new Non-Volatile memory technology promises to improve performance, efficiency that matches with ideal characteristics. The article reviews some of new non-volatile memory technologies. PADAM RAO

RAM, DRAM and Flash are three Landscape dominating memory technologies and With respect to ideal characteristics of memory each technology has some advantages technology described above, till now all known and disadvantages. The ideal memory S memory/storage technologies address only some technology should have low power of the characteristics. Static RAM (SRAM) is very consumption, high performance, high reliability, fast, but it's expensive, has low density and is not high density, low cost, and the ability to be used in any application. persistent. Dynamic RAM (DRAM) has better Besides computers, today’s portable electronics densities and is cheaper (but still expensive) at the have become computationally intensive devices as cost of being a little slower, and it's not persistent as the user interface has migrated to a fully well. Disks are cheap, highly dense and persistent, multimedia experience. To provide the but very slow. is between DRAM and performance required for these applications, the disk; it is a persistent solid-state memory with higher portable electronics designer uses multiple types of densities than DRAM, but its write latency is much memories: a medium-speed random access higher than the latter. memory for continuously changing , a high- Fortunately, it is possible to design a memory system speed memory for caching instructions to the CPU, that incorporates all these different technologies in and a slower, non-volatile memory for long-term a single . Such hierarchy allows information storage when the power is removed. the creation of a system that approaches the Combining all of these memory types into a single performance of the fastest component, the cost of memory has been a long-standing goal of the the cheapest component and energy consumption semiconductor industry. Today we watch the of the most power-efficient component. This is emergence of new non-volatile memory technologies, such as Phase-Change RAM (PCRAM), Magnetoresistive RAM (MRAM) and Resistive RAM RRAM), that promise to radically change the landscape of memory systems. The introduction of these new technologies will probably result in more complex memory hierarchies, but is likely to allow the construction of memory chips that are non-volatile, low-energy and have density and latency close to or better than current DRAM chips, improving performance/ efficiency and allowing memory systems to continue to scale up. Fig. 1 - Diagram illustrating the concept of memory hierarchy MEMORY

possible due to a phenomenon known as , so named because memory references Bitline tend to be localized in time and space. This can be Phase-change material so summarized: Contact Access device - If you use something once, you are likely to use it Insulator again (temporal locality). Wordline - If you use something once, you are likely to use Fig. 2 - Example of a phase-change . The current its neighbor (spatial locality) This phenomenon can is forced to pass through the phase-change material. Image be exploited by creating different memory levels; obtained from the first levels are smaller and more expensive, but amorphous, it is melt-quenched using a high- have fastest access, and the other layers are power electrical pulse that is abruptly cut of. To progressively bigger and cheaper, but have slower crystallize the material, it is heated above its access times. Appropriate heuristics are then crystallization temperature using a moderate power applied to decide which data will be accessed more pulse with a longer duration. Since the duration of often and place them at the first levels, and move this pulse varies according to the crystallization the data down on the hierarchy as it ceases to be speed of the material being used, this operation used frequently. Figure 1 depicts the concept of tends to dictate the writing speed of PCRAM. memory hierarchy. Reflectivity can vary up to 30%, but resistivity changes can be as large as five orders of Emerging Memory Technologies magnitude. There are several new Non-Volatile Memory (NVM) The principle of phase-change memory is known technologies under research. One study lists 13 since the 1960s, but only recent discoveries of such technologies: FRAM, MRAM, STTRAM, phase-change materials with faster crystallization PCRAM, NRAM, RRAM, CBRAM, SEM, Polymer, speeds led to the possibility of commercially Molecular, Racetrack, Holographic and Probe. feasible memory technology. The most important Most these technologies are in different stages of materials are chalcogenides such as Ge2Sb2Te5 maturity. Some of them are still in early research (GST), that can crystallize in less than 100 ns. stages, others have working prototypes, and some Figure 2 shows a memory cell based on phase- of them are already entering into commercial change principles. The SET operation is achieved manufacturing. by crystallizing the material and RESET by making it The article reviews three of these technologies: amorphous. Unlike Flash memory, PCRAM can be Phase-Change RAM, Resistive RAM (including switched from 0 to 1 and vice-versa without an ) and Magnetoresistive RAM (including ERASE operation. Spin-Torque Transfer RAM). All these fall into the Given the great difference in resistance, phase- category of the most actively researched today, are change materials can be easily used to store binary backed by solid companies of the technology states per cell (single-level cell) and even more industry, and considered most promising of being states (multi-level cell). commercially feasible. PCRAM is argued to be a scalable technology. As the feature density increases, phase change Phase-Change RAM (PCRAM) material layers become thinner and need less Phase-Change Random Access Memory (also current for the programming operations.

called PCRAM, PRAM or PCM) is currently the most It has been demonstrated to work in 20nm device

mature of the new memory technologies under prototype and is projected to scale down to 9nm. research. It relies on some materials, called phase- , change materials, that exist in two different phases PCRAM is argued to be a scalable with distinct properties: technology. It has been 1. An amorphous phase, characterized by high demonstrated to work in 20nm electrical resistivity; device prototype and is projected 2. A crystalline phase, characterized by low electrical resistivity. to scale down to 9nm. DRAM These two phases can be repeatedly and rapidly probably will not able to scale cycled by applying heat to the material. To make it down beyond 40nm., MEMORY

DRAM probably will not able to scale down beyond that voltage has been applied. When you turn off 40nm. Today PCRAM is positioned as a Flash the voltage, the remembers its most replacement. It offers great advantages over Flash, recent resistance until the next time you turn it on, but given the current limitations of access latency, whether that happens a day later or a year later. (...) energy consumption and endurance, further The ability to indefinitely store resistance values development is required in order to employ it as a means that a memristor can be used as a replacement for DRAM. nonvolatile memory.” This memristor device consisted of a crossbar of Resistive RAM (RRAM) platinum wires with titanium dioxide (TiO2) Despite the fact that PCRAM also uses resistance switches, as shown in Figure 4. Each switch consists variations to store values, the term Resistive RAM of a lower layer of perfect titanium dioxide (TiO2), (RRAM or ReRAM) has been applied to a distinct set which is electrically insulating, and an upper layer of technologies that explore the same of oxygen deficient titanium dioxide (TiO2 - x), phenomenon. Essentially these technologies fall which is conductive. The size of each layer can be into one of two categories: changed by applying voltage to the top electrode. If 1. Insulator resistive memories: based on bipolar a positive voltage is applied, the TiO2 – x ayer resistance switching properties of some metal thickness increases and the switch becomes oxides. The most important example is the conductive (ON state). A negative voltage has the memristor memory device, which will be further opposite effect (OFF state). This behavior matches described in more detail. the hysteresis loop previously shown in Figure 3. 2. Solid-electrolyte memories: based on solid- Titanium dioxide was used in the mentioned electrolyte containing mobile metal ions prototype, but several other oxides are known to sandwiched between a cathode and an anode. present similar bipolar resistive switching, and there Also known as Programmable Metallization Cell are multiple research projects in motion to explore (PMC) or Conductive Bridge RAM (CBRAM). these other materials for similar memory device There is a long list of RRAM technologies. This article will concentrate attention on the memristor, which is currently the most promising RRAM 1.0

technology under research. 0.5

Memristor t n e r

Since the 19th century, three fundamental passive r 0.0 u circuit elements were known: the resistor, the c inductor and the . In 1971, Leon Chua -0.5 theorized the existence of a fourth passive circuit element, which he called the memristor, but no -1.0 actual physical device with memristive properties -1.0 -0.5 0.0 0.5 1.0

could be constructed. In 2008, a group of scientists voltage reported the invention of a device that behaved as Fig. 3 - Relationship between current and voltage in a predicted for a memristor. Later the same year, an memristor, known as a ‘hysteresis loop’. Extracted from Wang, article detailed how that device could be used to 2008 create nonvolatile memories. The property of memristors particularly relevant to memory devices is the nonlinear relationship between current (I) and voltage (V), depicted on Figure 3. In the words of R. Stanley Williams, the leader of the research team that invented the first memristor device: “Memristor is a contraction of 'memory resistor', because that is exactly its function: to remember its history. A memristor is a two-terminal device whose resistance depends on the magnitude and polarity of the voltage applied to it and the length of time Fig. 4. - How memristance works

MEMORY implementations. , Memristive memory technology is less mature than MRAM is a relatively mature PCRAM, and DIMM prototypes aren't available yet, technology, but current but there are already a few that predict implementations suffer from some fundamental characteristics of such density and energy constraints memories. that seriously affect its capacity Scalability is one aspect where memristor is most to compete with existing memory efficient. A cell size of 10 nm has been achieved technologies such as DRAM or and a density between 4-5 nm is predicted for the Flash. next few years. , Beyond that, memristor memory devices can development called STT-RAM has a good chance benefit from multi-level cell (MLC) storage and 3D to overcome these problems and position MRAM die stacking, which can greatly improve the overall as a commercially feasible NVM alternative. density. STT-RAM Conventional MRAM (also called “toggle-mode” Magnetoresistive RAM (MRAM) MRAM) uses a current induced magnetic field to Magnetoresistive RAM (MRAM), sometimes called switch the MTJ magnetization. The amplitude of the Magnetic RAM, is a memory technology that magnetic field must increase as the size of MTJ explores a component called Magnetic Tunnel scales, which compromises MRAM scalability. Spin- Junction (MTJ). Torque Transfer MRAM (STT-RAM) technology tries An MTJ, depicted on Figure 5, consists of two to achieve better scalability by employing a ferromagnetic layers separated by an oxide tunnel different write mechanism based on spin barrier layer (e.g.,: MgO). One of the polarization. ferromagnetic layers, called the reference layer, Read/write latencies for STT-RAM are very low. One keeps its magnetic direction fixed, while the other, study estimates a read latency of 5.5 ns, faster than called the free layer, can have its direction changed DRAM. The same study predicts slower write by means of either a magnetic field or a polarized speeds, around 12.5 ns, but they are still very fast. current. When both the reference layer and the free Endurance is excellent, reaching figures above layer have the same direction, the resistance of the 1015cycles. MTJ is low. If they have different directions, the The less favorable aspect of STT-RAM is density. Its resistance is high. This phenomenon is known as cells have less density than currently achieved by Tunneling Magneto-Resistance (TMR). DRAM. MRAM is a relatively mature technology, but current From one research article proposing that MRAM implementations suffer from density and energy constraints that seriously affect its capacity to could be the “” technology that compete with existing memory technologies such would render obsolete all other components of as DRAM or Flash. An specific type of MRAM under memory hierarchy (SRAM, DRAM, disks, _ash). Since then, papers on MRAM (including STTRAM) B B often mention its possible future as “universal memory”. In practice, due to MRAM current density limitations, it is not seen as a real replacement for Reference Layer Reference Layer DRAM as main memory (not to mention disks). Current ideas focus on using STT-RAM as an SRAM MgO MgO replacement for building memories, due to

Free Layer Free Layer its excellent read/write speeds, endurance, and much superior energy efficiency.

A A Comparison between Memory (a) (b) Technologies So far, the article presented the most promising Fig. 5 - A conceptual view of MTJ structure. (a) Anti-parallel NVM technologies. To compare each one of them (high resistance), which indicates “1” state; (b) Parallel (low resistance), which indicates “0” state. as well as against the main current memory MEMORY

Table 1 - Comparison of memory technologies. The colors indicate how well each technology scores against the specified criteria. Green means a good score, red means a bad score, yellow is for scores in between. hierarchy components: SRAM, DRAM, disks and memory for persistent storage. So far, this design flash. In order to do this comparison, the following has enabled growth according to Moore's Law. are the set of criteria: However, limitations of these technologies threaten 1. Maturity: whether the technology is currently the ability to sustain this growth rate during the next used in the market or it is in early or later research decade, creating a “power wall”. In order to allow stages before being commercially mature. continuing growth, it will be necessary to develop 2. Cell size: the cell size, using standard feature size new memory technologies. (F2). As consequence, today we watch the emergence of 3. Read Latency: the speed for reading values from new memory technologies that promise to change a memory cell. radically the landscape of 4. Write Latency: the speed for writing values to a systems. The ideal future memory technology memory cell. should be non-volatile, low-cost, highly dense, 5. Endurance: the number of write cycles that a energy efficient, fast and with high endurance. Such memory cell endures before eventually wearing ideal technology would be a “universal memory”, out. eliminating the need for complex hierarchies in the 6. Energy: energy spent per bit access. Related to memory subsystem. dynamic power. Unfortunately, as we've seen, none of the presented 7. Static Power: whether power needs to be spent technologies has today all the necessary attributes while not accessing the memory device. This to accomplish that. PCRAM is not fast enough and includes refreshing solid memory contents due to does not have the necessary endurance. STT- energy leakage or keeping disks spinning to MRAM cannot achieve high densities. RRAM has achieve lower access latencies. very limited endurance. Until these issues are 8. Non-volatility: whether the memory technology addressed, there will be no single “universal is volatile or not. memory”. Actually, the introduction of these new Based on the data presented in the last sections of technologies will probably result in more complex this work, table 1 compares the discussed memory hierarchies instead of simpler ones. These technologies against the defined set of criteria. technologies can allow the construction of memory chips that are non-volatile, low-energy, highly Conclusion dense and with latency close to or better than For almost 30 years, computer memory systems current DRAM chips. have been composed of SRAM for caches, DRAM for main memory and magnetic disks and flash