In this paper, we show the feasibility of low supply voltage for SRAM (Static Random Access Memory) by adding error correction coding (ECC). In SRAM, the memory matrix needs to be powered for data retentive standby operation, resulting in standby leakage current. Particularly for low duty- cycle systems, the energy consumed due to standby leakage current can become significant. Lowering the supply voltage (VDD) during standby mode to below the specified data retention voltage (DRV) helps decrease the leakage current. At these VDD levels errors start to appear, which we can remedy by adding ECC. We show in this paper that addition of a simple single error correcting (SEC) ECC enables us to decrease the leakage current by 45% and leakage power by 72%. We verify this on a large set of commercially available standard 40nm SRAMs.
To continue reducing voltage in scaled technologies, both circuit and architecture-level resiliency techniques are needed to tolerate process-induced defects, variation, and aging in SRAM cells. Many different resiliency schemes have been proposed and evaluated, but most prior results focus on voltage reduction instead of energy reduction. At the circuit level, device cell architectures and assist techniques have been shown to lower Vmin for SRAM, while at the architecture level, redundancy and cache disable techniques have been used to improve resiliency at low voltages. This paper presents a unified study of error tolerance for both circuit and architecture techniques and estimates their area and energy overheads. Optimal techniques are selected by evaluating both the error-correcting abilities at low supplies and the overheads of each technique in a 28nm. The results can be applied to many of the emerging memory technologies.
Emerging Memories (EMs) could benefit from Error Correcting Codes (ECCs) able to correct few errors in a few nanoseconds. The low latency is necessary to meet the DRAM- like and/or eXecuted-in-Place requirements of Storage Class Memory devices. The error correction capability would help manufacturers to cope with unknown failure mechanisms and to fulfill the market demand for a rapid increase in density. This paper shows the design of an ECC decoder for a shortened BCH code with 256-data-bit page able to correct three errors in less than 3 ns. The tight latency constraint is met by pre-computing the coefficients of carefully chosen Error Locator Polynomials, by optimizing the operations in the Galois Fields and by resorting to a fully parallel combinatorial implementation of the decoder. The latency and the area occupancy are first estimated by the number of elementary gates to traverse, and by the total number of elementary gates of the decoder. Eventually, the implementation of the solution by Synopsys topographical synthesis methodology in 54nm logic gate length CMOS technology gives a latency lower than 3 ns and a total area less than \(250 \cdot 10^3 \mu m^2\).
Magnetic spin-based memory technologies are a promising solution to overcome the incoming limits of microelectronics. Nevertheless, the long write latency and high write energy of these memory technologies compared to SRAM make it difficult to use these for fast microprocessor memories, such as L1- Caches. However, the recent advent of the Spin Orbit Torque (SOT) technology changed the story: indeed, it potentially offers a writing speed comparable to SRAM with a much better density as SRAM and an infinite endurance, paving the way to a new paradigm in processor architectures, with introduction of non- volatility in all the levels of the memory hierarchy towards full normally-off and instant-on processors. This paper presents a full design flow, from device to system, allowing to evaluate the potential of SOT for microprocessor cache memories and very encouraging simulation results using this framework.
The capacity of embedded memory on LSIs has kept increasing. It is important to reduce the leakage power of embedded memory for low-power LSIs. In fact, the ITRS predicts that the leakage power in embedded memory will account for 40% of all power consumption by 2024 . A spin transfer torque magneto-resistance random access memory (STT-MRAM) is promising for use as non-volatile memory to reduce the leakage power. It is useful because it can function at low voltages and has a lifetime of over 1016 write cycles . In addition, the STT-MRAM technology has a smaller bit cell than an SRAM. Making the STT-MRAM is suitable for use in high-density products [3–7]. The STT-MRAM uses magnetic tunnel junction (MTJ). The MTJ has two states: a parallel state and an anti-parallel state. These states mean that the magnetization direction of the MTJ’s layers are the same or different. The directions pair determines the MTJ’s magneto- resistance value. The states of MTJ can be changed by the current flowing. The MTJ resistance becomes low in the parallel state and high in the anti-parallel state. The MTJ potentially operates at less than 0.4 V . In other hands, it is difficult to design peripheral circuitry for an STT-MRAM array at such a low voltage. In this paper, we propose a counter-based read circuit that functions at 0.4 V, which is tolerant of process variation and temperature fluctuation.
3D integration of solid-state memories and logic, as demonstrated by the Hybrid Memory Cube (HMC), offers major opportunities for revisiting near-memory computation and gives new hope to mitigate the power and performance losses caused by the “memory wall”. In this paper we present the first exploration steps towards design of the Smart Memory Cube (SMC), a new Processor-in-Memory (PIM) architecture that enhances the capabilities of the logic-base (LoB) in HMC. An accurate simulation environment has been developed, along with a full featured software stack. All offloading and dynamic overheads caused by the operating system, cache coherence, and memory management are considered, as well. Benchmarking results demonstrate up to 2X performance improvement in comparison with the host SoC, and around 1.5X against a similar host-side accelerator. Moreover, by scaling down the voltage and frequency of PIM’s processor it is possible to reduce energy by around 70% and 55% in comparison with the host and the accelerator, respectively.
Multiple-channel die-stacked DRAMs have been used for maximizing the performance and minimizing the power of memory access in 2.5D/3D system chips. Stacked DRAM dies can be used as a cache for the processor die in 2.5D/3D system chips. Typically, modern processor system-on-chips (SOCs) have three-level caches, L1, L2, and L3. Could the DRAM cache be used to replace which level of caches? In this paper, we derive an inequality which can aid the designer to check if the designed DRAM cache can provide better performance than the L3 cache. Also, design considerations of DRAM caches for meet the inequality are discussed. We find that a dilemma of the DRAM cache access time and associativity exists for providing better performance than the L3 cache. Organizing multiple channels into a DRAM cache is proposed to cope with the dilemma.
This paper briefly discusses a new architecture, Computation-In-Memory (CIM Architecture), which performs “processing-in-memory”. It is based on the integration of storage and computation in the same physical location (crossbar topology) and the use of non-volatile resistive-switching technology (memristive devices or memristors in short) instead of CMOS technology. The architecture has the potential of improving the energy-delay product, computing efficiency and performance area by at least two orders of magnitude.
Three-dimensional (3D) integration using through- silicon via (TSV) has been used for memory designs. Content addressable memory (CAM) is an important component in digital systems. In this paper, we propose an evaluation tool for 3D CAMs, which can aid the designer to explore the delay and power of various partitioning strategies. Delay, power, and energy models of 3D CAM with respect to different architectures are built as well.
The energy efficiency of today’s microcontrollers is supported by the extensive usage of low-power mechanisms. A full power-down requires in many cases a complex, and maybe error prone, administration scheme, because data from the volatile memory have to be stored in a flash based back- up memory. New types of non-volatile memory, e.g. in RRAM technology, are faster and consumes a fraction of the energy compared to flash technology. This paper evaluates power gating for WSN with RRAM as back-up memory.