On-chip SRAM Research Articles

Analyzing the Impact of Joint Optimization of Cell Size, Redundancy, and ECC on Low-Voltage SRAM Array Total Area

The increasing power consumption of processors has made power reduction a first-order priority in processor design. Voltage scaling is one of the most powerful power-reduction techniques introduced to date, but is limited to some minimum voltage <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> . Below <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> on-chip SRAM cells cannot all operate reliably due to increased process variability with technology scaling. The use of larger SRAM cells, which are less sensitive to process variability, allows a reduction in <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> . However, since the large-scale memory structures such as last-level caches (LLCs) often determine the <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> of processors, these structures cannot afford to use large SRAM cells due to the resulting increase in die area. In this paper we first propose a joint optimization of LLC cell size, the number of redundant cells, and the strength of error-correction coding (ECC) to minimize total SRAM area while meeting yield and <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> targets. The joint use of redundant cells and ECC enables the use of smaller cell sizes while maintaining design targets. Smaller cell sizes more than make up for the extra cells required by redundancy and ECC. In 32-nm technology our joint approach yields a 27% reduction in total SRAM area (including the extra cells) when targeting 90% yield and 600 mV <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DDMIN</sub> . Second, we demonstrate that the ECC used to repair defective cells can be combined with a simple architectural technique, which can also fix particle-induced soft errors, without increasing ECC strength or processor runtime.

Minimizing Energy Consumption of Embedded Systems via Optimal Code Layout

Code repositioning is a well-known method of reducing inefficient off-chip memory accesses by streamlining cache behavior. Embedded systems with predetermined applications can achieve further improvement with the addition of fast and energy efficient scratchpad memory (SPM) on chip and moving frequent accesses code and/or data from main memory to SPM. While many researchers have attempted to either streamline cache accesses or improve the effectiveness of SPM, few studies focus on exploring their joint synergy. This study proposes integer linear programming (ILP) models that include both code repositioning and SPM code selection to identify the optimal code layout and reduce energy consumption in embedded systems with a cache and SPM. This study also proposes a two-stage metaheuristic algorithm. Experimental results reveal that 1) allocating a dedicated portion of the on-chip SRAM to the SPM is not always better than using a cache-only configuration and 2) it is not trivial to select code objects for the SPM. As much as 55 percent additional energy can be saved by applying both code repositioning and SPM code selection techniques.

Concise Lookup Tables for IPv4 and IPv6 Longest Prefix Matching in Scalable Routers

We present a distinct longest prefix matching (LPM) lookup scheme able to achieve exceedingly concise lookup tables (CoLT), suitable for scalable routers. Based on unified hash tables for handling both IPv4 and IPv6 simultaneously, CoLT excels over previous mechanisms in: 1) lower on-chip storage for lookup tables; 2) simpler table formats to enjoy richer prefix aggregation and easier implementation; and 3) most importantly, deemed the only design able to accommodate both IPv4 and IPv6 addresses uniformly and effectively. As its hash tables permit multiple possible buckets to hold each prefix (following a migration rule to avoid false positives altogether), CoLT exhibits the best memory efficiency and can launch parallel search over tables during every LPM lookup, involving fewer cycles per lookup when on-chip memory is used to implement hash tables. With 16 (or 32) on-chip SRAM blocks clocked at 500 MHz (achievable in today's 65-nm technology), it takes 2 (or 1.6) cycles on average to complete a lookup, yielding 250 (or 310+) millions of packets per second (MPPS) mean throughput. Being hash-oriented, CoLT well supports incremental table updates, besides its high table utilization and lookup throughput.

A reduced memory bandwidth and high throughput HDTV motion compensation decoder for H.264/AVC High 4:2:2 profile

This article presents the HP422-MoCHA: optimized Motion Compensation hardware architecture for the High 4:2:2 profile of H.264/AVC video coding standard. The proposed design focuses on real-time decoding for HDTV 1080p (1,920 × 1,080 pixels) at 30 fps. It supports multiple sample bit-width (8, 9, or 10 bits) and multiple chroma sub-sampling formats (4:0:0, 4:2:0, and 4:2:2) to provide enhanced video quality experience. The architecture includes an optimized sample interpolator that processes luma and chroma samples in two parallel datapaths and features quarter sample accuracy, bi-prediction and weighted prediction. HP422-MoCHA also includes a hardwired Motion Vector Predictor, supporting temporal and spatial direct predictions. A novel memory hierarchy implemented as a 3-D Cache reduces the frame memory access, providing, on average, 62% of bandwidth and 80% of clock cycles reduction. The design was implemented in a Xilinx Virtex-II PRO FPGA, and also in an ASIC with a TSMC 0.18 μm standard cells technology. The ASIC implementation occupies 102 K equivalent gates and 56.5 KB of on-chip SRAM in a 3.8 × 3.4 mm2 area. It presents a power consumption of 130 mW. Both implementations reach a maximum operation frequency of ~100 MHz, being able to motion compensate 37 bi-predictive frames or 69 predictive fps. The minimum required frequency to ensure the real-time decoding for HD1080p at 30 fps is 82 MHz. Since HP422-MoCHA is the first Motion Compensation architecture for the High 4:2:2 profile found in the literature, a Main profile MoCHA was used for comparison purposes, showing the highest throughput among all presented works. However, the HP422-MoCHA architecture also reaches the highest throughput when compared with the other published Main profile MC solutions, even considering the significantly higher complexity of the High 4:2:2 profile.

Cache-Integrated Network Interfaces: Flexible On-Chip Communication and Synchronization for Large-Scale CMPs

Per-core scratchpad memories (or local stores) allow direct inter-core communication, with latency and energy advantages over coherent cache-based communication, especially as CMP architectures become more distributed. We have designed cache-integrated network interfaces, appropriate for scalable multicores, that combine the best of two worlds – the flexibility of caches and the efficiency of scratchpad memories: on-chip SRAM is configurably shared among caching, scratchpad, and virtualized network interface (NI) functions. This paper presents our architecture, which provides local and remote scratchpad access, to either individual words or multiword blocks through RDMA copy. Furthermore, we introduce event responses, as a technique that enables software configurable communication and synchronization primitives. We present three event response mechanisms that expose NI functionality to software, for multiword transfer initiation, completion notifications for software selected sets of arbitrary size transfers, and multi-party synchronization queues. We implemented these mechanisms in a four-core FPGA prototype, and measure the logic overhead over a cache-only design for basic NI functionality to be less than 20%. We also evaluate the on-chip communication performance on the prototype, as well as the performance of synchronization functions with simulation of CMPs with up to 128 cores. We demonstrate efficient synchronization, low-overhead communication, and amortized-overhead bulk transfers, which allow parallelization gains for fine-grain tasks, and efficient exploitation of the hardware bandwidth.

Design of Highly Efficient Self-Adaptive H.264 Loop Filter

A highly efficient, single-port SRAM-based loop filter based on H.264 is presented. Hybrid edge filtering order, efficient memory organization for both on-chip SRAM and transposition buffers, and five-stage pipeline architect are adopted in the design. Efficient transposition buffer organization reduces the gate counts and the memory access times. By using hybrid edge filtering order and alternating output memory update scenario, zero stall cycle in pipeline flow is realized. To balance on-chip memory size and bus bandwidth, the design can be configured to operate in on-chip memory and off-chip memory modes.

Enhancing Counting Bloom Filters Through Huffman-Coded Multilayer Structures

Bloom Filters are efficient randomized data structures for membership queries on a set with a certain known false positive probability. Counting Bloom Filters (CBFs) allow the same operation on dynamic sets that can be updated via insertions and deletions with larger memory requirements. This paper first presents a simple tight upper bound for counters overflow probability in CBFs, which is adopted in the design of more efficient CBFs. On the basis of such theoretical achievements, we introduce the idea of a hierarchical structure as well as the use of Huffman code to improve standard CBFs in terms of fast access and limited memory consumption (up to 50% of memory saving). The target could be the implementation of the compressed data structures in the small (but fast) local memory or on-chip SRAM of devices such as network processors. As an application of our algorithms, an anti-evasion system is finally proposed.

Low-cost sound-based localization using programmable mixed-signal systems-on-chip

There have been extensive theoretical studies on sound-based localization using both, a pair of microphones and microphone arrays. In contrast, there has been much less work on implementing and experimenting sound-based localization realized as customized electronic designs. This paper presents a low-cost implementation of the sound-based localization method proposed in Halupka et al. [11]. A new method called wave counting is proposed in this paper, as an alternative to the Maximum Likelihood procedure used in [11]. The implementation uses PSoC programmable mixed-signal embedded system-on-chip, which incorporates microcontroller, on-chip SRAM and flash memory, programmable digital blocks, and programmable analog blocks, all integrated on the same chip. The paper presents an extensive set of experiments to characterize the quality of localization using the proposed low-cost design.

A Hardware-Efficient Multi-Resolution Block Matching Algorithm and its VLSI Architecture for High Definition MPEG-Like Video Encoders

High throughput, heavy bandwidth requirement, huge on-chip memory consumption, and complex data flow control are major challenges in high definition integer motion estimation hardware implementation. This paper proposes an efficient very large scale integration architecture for integer multi-resolution motion estimation based on optimized algorithm. There are three major contributions in this paper. First, this paper proposes a hardware friendly multi-resolution motion estimation algorithm well-suited for high definition video encoder. Second, parallel processing element (PE) array structure is proposed to implement three-level hierarchical motion estimation, only 256PEs are enough for one reference frame real-time high definition motion estimation by efficient PE reuse. Third, efficient on-chip reference pixel buffer sharing mechanism between integer and fractional motion estimation is proposed with almost 50% SRAM saving and memory bandwidth reduction. The proposed multi-resolution motion estimation algorithm reached a good balance between complexity and performance with rate distortion optimized variable block size motion estimation support. Also, we have achieved moderate logic circuit and on-chip SRAM consumption. The proposed architecture is well-suited for all MPEG-like video coding standards such as H.264, audio video coding standard, and VC-1.

A Differential Data-Aware Power-Supplied (D$^{2}$AP) 8T SRAM Cell With Expanded Write/Read Stabilities for Lower VDDmin Applications

Due to global and local process variations, on-chip SRAM suffers failures at a low supply voltage (VDD). This study proposes a differential data-aware power-supplied D <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> AP 8T SRAM cell to address the stability and trade-off-issues between write and half-select accesses that still remain in the conventional 8T and 6T cells. Powered by its bitline pair, the proposed 8T cell applies differential data-aware-supplied voltages to its cross-coupled inverters to increase both stability margins for write and half-select accesses. A boosted bitline scheme also improves the read cell current. Two 39 Kb SRAM macros, D <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> AP-8T and conventional 8T, with the same peripheral circuits were fabricated on the same testchip with 45 nm and 40 nm processes. The measured VDDmin for the D <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> AP-8T macro is 240 mV-200 mV lower than that of the conventional 8T macro across lots, wafers and dies.

Dynamic Multiway Segment Tree for IP Lookups and the Fast Pipelined Search Engine

A dynamic multiway segment tree (DMST) is proposed for IP lookups in this paper. DMST is designed for dynamic routing tables that can dynamically insert and delete prefixes. DMST is implemented as a B-tree that has all distinct end points of ranges as its keys. The complexities of search, insertion, deletion, and memory requirement are the same as the existing multiway range tree (MRT) and prefix in B-tree (PIBT) for prefixes. In addition, a pipelined DMST search engine is proposed to further speed up the search operations. The proposed pipelined DMST search engine uses off-chip SRAMs instead of on-chip SRAMs because the capacity of the latter is too small to hold large routing tables and the cost of the latter is too expensive. By utilizing current FPGA and off-chip SRAM technologies, our proposed five-stage pipelined search engine can achieve the worst case throughputs of 33.3 and 41.7 million packets per second (Mpps) with 144-bit and 288-bit wide SRAM blocks, respectively. Furthermore, a straightforward extension of the pipelined search engine with multiple independent off-chip SRAMs can achieve the throughput of 200 Mpps which is equivalent to 102 Gbps for minimal Ethernet packets of size 64 bytes.

SUSE: Superior Storage-Efficiency for Routing Tables Through Prefix Transformation and Aggregation

A novel storage design for IP routing table construction is introduced on the basis of a single set-associative hash table to support fast longest prefix matching (LPM). The proposed design involves two key techniques to lower table storage required drastically: 1) storing transformed prefix representations; and 2) accommodating multiple prefixes per table entry via prefix aggregation, achieving superior storage-efficiency (SUSE). With each prefix (p(x)) maneuvered as a polynomial, p(x) - q(x) x g(a;) + r(x) based on a divisor g(x), SUSE keeps only q(x) rather than full and long p(x) in an r(x)-indexed table with 2degree(g(x)) entries, because q(x) and r(x) uniquely identify p(x). Additionally, using r(x) as the hash index exhibits better distribution than do original prefixes, reducing hash collisions, which can be tolerated further by the set-associative design. Given a set of chosen prefix lengths (called treads), all prefixes are rounded down to nearest treads under SUSE before hashed to the table using their transformed representations so that prefix aggregation opportunities abound in hash entries. SUSE yields significant table storage reduction and enjoys fast lookups and speedy incremental updates not possible for a typical trie-based design, with the worst-case lookup time shown upper-bounded theoretically by the number of treads (?) but found experimentally to be 4 memory accesses when ? equals 8. SUSE makes it possible to fit a large routing table with 256 K (or even 1 M) prefixes in on-chip SRAM by today's ASIC technology. It solves both the memory- and the bandwidth-intensive problems faced by IP routing.

Technologies for Ultradynamic Voltage Scaling

Energy efficiency of electronic circuits is a critical concern in a wide range of applications from mobile multi-media to biomedical monitoring. An added challenge is that many of these applications have dynamic workloads. To reduce the energy consumption under these variable computation requirements, the underlying circuits must function efficiently over a wide range of supply voltages. This paper presents voltage-scalable circuits such as logic cells, SRAMs, ADCs, and dc-dc converters. Using these circuits as building blocks, two different applications are highlighted. First, we describe an H.264/AVC video decoder that efficiently scales between QCIF and 1080p resolutions, using a supply voltage varying from 0.5 V to 0.85 V. Second, we describe a 0.3 V 16-bit micro-controller with on-chip SRAM, where the supply voltage is generated efficiently by an integrated dc-dc converter.

A 212 MPixels/s 4096 $\times$ 2160p Multiview Video Encoder Chip for 3D/Quad Full HDTV Applications

Multiview video coding (MVC) plays an important role in a 3-D video system. In addition, the resolution of HDTV is increasing to present more vivid perception for users. To realize real-time processing of dozens of TOPS, VLSI solution is necessary. However, ultra high computational complexity, a large amount of external memory bandwidth and on-chip SRAM size, and complex MVC prediction structures are three main design challenges of implementation of MVC hardware architecture. In this paper, an MVC single-chip encoder is proposed for H.264/AVC Multiview High Profile and High Profile for 3-D and quad full high definition (QFHD) TV applications, respectively. The 4096 × 2160 p multiview video encoder chip is implemented on a 11.46 mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> die with 90 nm CMOS technology. An eight-stage macroblock pipelined architecture with proposed system scheduling and cache-based prediction core supports real-time processing from one-view 4096 × 2160 p to seven-view 720 p videos. The 212 Mpixels/s throughput is 3.4 to 7.7 times higher than previous work. The 407 Mpixels/W power efficiency is achieved, and 94% on-chip SRAM size and 79% external memory bandwidth are saved by the proposed techniques.

CMOS Ultrasound Transceiver Chip for High-Resolution Ultrasonic Imaging Systems

The proposed CMOS ultrasound transceiver chip will enable the development of portable high resolution, high-frequency ultrasonic imaging systems. The transceiver chip is designed for close-coupled MEMS transducer arrays which operate with a 3.3-V power supply. In addition, a transmit digital beamforming system architecture is supported in this work. A prototype chip containing 16 receive and transmit channels with preamplifiers, time-gain compensation amplifiers, a multiplexed analog-to-digital converter with 3 kB of on-chip SRAM, and 50-MHz resolution time delayed excitation pulse generators has been fabricated. By utilizing a shared A/D converter architecture, the number of A/D converter and SRAM is cut down to one, unlike typical digital beamforming systems which need 16 A/D converters for 16 receive channels. The chip was fabricated in a 0.35-mum standard CMOS process. The chip size is 10 mm(2), and its average power consumption in receive mode is approximately 270 mW with a 3.3-V power supply. The transceiver chip specifications and designs are described, as well as measured results of each transceiver component and initial pulse-echo experimental results are presented.

Compiler-directed scratchpad memory management via graph coloring

Scratchpad memory (SPM), a fast on-chip SRAM managed by software, is widely used in embedded systems. This article introduces a general-purpose compiler approach, called memory coloring, to assign static data aggregates, such as arrays and structs, in a program to an SPM. The novelty of this approach lies in partitioning the SPM into a pseudo--register file (with interchangeable and aliased registers), splitting the live ranges of data aggregates to create potential data transfer statements between SPM and off-chip memory, and finally, adapting an existing graph coloring algorithm for register allocation to assign the data aggregates to the pseudo--register file. Our experimental results using a set of 10 C benchmarks from MediaBench and MiBench show that our methodology is capable of managing SPMs efficiently and effectively for large embedded applications. In addition, our SPM allocator can obtain close to optimal solutions when evaluated and compared against an existing heuristics-based SPM allocator and an ILP-based SPM allocator.

A 201.4 GOPS 496 mW Real-Time Multi-Object Recognition Processor With Bio-Inspired Neural Perception Engine

A 201.4 GOPS real-time multi-object recognition processor is presented with a three-stage pipelined architecture. Visual perception based multi-object recognition algorithm is applied to give multiple attentions to multiple objects in the input image. For human-like multi-object perception, a neural perception engine is proposed with biologically inspired neural networks and fuzzy logic circuits. In the proposed hardware architecture, three recognition tasks (visual perception, descriptor generation, and object decision) are directly mapped to the neural perception engine, 16 SIMD processors including 128 processing elements, and decision processor, respectively, and executed in the pipeline to maximize throughput of the object recognition. For efficient task pipelining, proposed task/power manager balances the execution times of the three stages based on intelligent workload estimations. In addition, a 118.4 GB/s multi-casting network-on-chip is proposed for communication architecture with incorporating overall 21 IP blocks. For low-power object recognition, workload-aware dynamic power management is performed in chip-level. The 49 mm2 chip is fabricated in a 0.13 ?m 8-metal CMOS process and contains 3.7 M gates and 396 KB on-chip SRAM. It achieves 60 frame/sec multi-object recognition up to 10 different objects for VGA (640 × 480) video input while dissipating 496 mW at 1.2 V. The obtained 8.2 mJ/frame energy efficiency is 3.2 times higher than the state-of-the-art recognition processor.

Energy and Performance Optimization of Demand Paging With OneNAND Flash

New fusion memory devices consisting of multiple heterogeneous memory components in a single die or package offer efficient ways to optimize embedded systems in terms of energy, performance, and cost. Samsung Electronics recently announced the OneNAND fusion memory, in which a NAND flash array is integrated with dual SRAM buffers to provide a nor-type I/O interface. OneNAND has the low cost and large capacity of a NAND flash but also permits eXecution-in-Place (XIP) like a nor flash. The deployment of such devices requires careful system-level resource management because of their impact on energy consumption and performance, and existing memory optimization techniques, such as the demand paging used with NAND flash, may no longer be appropriate for systems with a fusion memory. We introduce a new online demand paging scheme that fully exploits the XIP capability of OneNAND flash by classifying pages as load preferred (residing in the on-chip SRAM) and XIP preferred (accessed directly from the OneNAND flash and discarded after use). This achieves, on average, a 26% reduction in energy consumption and a 19% increase in performance, compared with conventional NAND flash demand paging.

A Proactive Wearout Recovery Approach for Exploiting Microarchitectural Redundancy to Extend Cache SRAM Lifetime

Microarchitectural redundancy has been proposed as a means of improving chip lifetime reliability. It is typically used in a reactive way, allowing chips to maintain operability in the presence of failures by detecting and isolating, correcting, and/or replacing components on a first-come, first-served basis only after they become faulty. In this paper, we explore an alternative, more preferred method of exploiting microarchitectural redundancy to enhance chip lifetime reliability. In our proposed approach, redundancy is used proactively to allow non-faulty microarchitecture components to be temporarily deactivated, on a rotating basis, to suspend and/or recover from certain wearout effects. This approach improves chip lifetime reliability by warding off the onset of wearout failures as opposed to reacting to them posteriorly. Applied to on-chip cache SRAM for combating NBTI-induced wearout failure, our proactive wearout recovery approach increases lifetime reliability (measured in mean-time-to-failure) of the cache by about a factor of seven relative to no use of microarchitectural redundancy and a factor of five relative to conventional reactive use of redundancy having similar area overhead.

A Five-Stage Pipeline, 204 Cycles/MB, Single-Port SRAM-Based Deblocking Filter for H.264/AVC

This paper describes the design and VLSI implementation of a highly efficient, single-port SRAM-based deblocking filter. It can achieve 204 cycles/macroblock throughput for H.264/AVC real-time decoding. Several deblocking filter designs in the literature have been compared and the possibility of realizing them in a pipeline is studied. Eventually we came up with a completely new design which has a five-stage pipeline with gated clock to increase system throughput while reducing power. Data hazards and structure hazards, which are the two most critical issues for a pipelined filter, are analyzed and resolved. Efficient memory organization for both on-chip SRAM and transposition buffers is employed. By using innovative hybrid edge filtering sequence and out-of-order memory update scenario, we obtain zero stall cycle in normal pipeline flow, making the best out of a pipelined architecture. Compared with existing designs, our design achieves at least 18% clock cycle reduction, as well as 20% lower power consumption owing to its efficient pipeline and memory architecture. The total gate count is comparable to other designs in literature without using any expensive two-port or dual-port on-chip SRAMs.