Tetris:A Heuristic Static Memory Management Framework for Uniform Memory Multicore Neural Network Accelerators

Uniform memory multicore neural network accelerators(UNNAs)furnish huge computing power to emerging neural network applications.Meanwhile,with neural network architectures going deeper and wider,the limited memory capacity has become a constraint to deploy models on UNNA platforms.Therefore how to efficiently manage memory space and how to reduce workload footprints are urgently significant.In this paper,we propose Tetris:a heuristic static memory management framework for UNNA platforms.Tetris reconstructs execution flows and synchronization relationships among cores to analyze each tensor’s liveness interval.Then the memory management problem is converted to a sequence permutation problem.Tetris uses a genetic algorithm to explore the permutation space to optimize the memory management strategy and reduce memory footprints.We evaluate several typical neural networks and the experimental results demonstrate that Tetris outperforms the state-of-the-art memory allocation methods,and achieves an average memory reduction ratio of 91.9%and 87.9%for a quad-core and a 16-core Cambricon-X platform,respectively.

Recovery of single event upset in advanced complementary metal-oxide semiconductor static random access memory cells

Using computer-aided design three-dimensional （3D） simulation technology, the recovery mechanism of single event upset and the effects of spacing and hit angle on the recovery are studied. It is found that the multi-node charge collection plays a key role in recovery and shielding the charge sharing by adding guard rings. It cannot exhibit the recovery effect. It is also indicated that the upset linear energy transfer （LET） threshold is kept constant while the recovery LET threshold increases as the spacing increases. Additionally, the effect of incident angle on recovery is analysed and it is shown that a larger angle can bring about a stronger charge sharing effect, thus strengthening the recovery ability.

Synergistic effects of total ionizing dose on single event upset sensitivity in static random access memory under proton irradiation

Synergistic effects of the total ionizing dose （TID） on the single event upset （SEU） sensitivity in static random access memories （SRAMs） were studied by using protons. The total dose was cumulated with high flux protons during the TID exposure, and the SEU cross section was tested with low flux protons at several cumulated dose steps. Because of the radiation-induced off-state leakage current increase of the CMOS transistors, the noise margin became asymmetric and the memory imprint effect was observed.

A review on SRAM-based computing in-memory:Circuits,functions,and applications

Artificial intelligence(AI)processes data-centric applications with minimal effort.However,it poses new challenges to system design in terms of computational speed and energy efficiency.The traditional von Neumann architecture cannot meet the requirements of heavily datacentric applications due to the separation of computation and storage.The emergence of computing inmemory(CIM)is significant in circumventing the von Neumann bottleneck.A commercialized memory architecture,static random-access memory(SRAM),is fast and robust,consumes less power,and is compatible with state-of-the-art technology.This study investigates the research progress of SRAM-based CIM technology in three levels:circuit,function,and application.It also outlines the problems,challenges,and prospects of SRAM-based CIM macros.

Pattern imprinting in deep sub-micron static random access memories induced by total dose irradiation

Pattem imprinting in deep sub-micron static random access memories （SRAMs） during total dose irradiation is inves- tigated in detail. As the dose accumulates, the data pattern of memory cells loading during irradiation is gradually imprinted on their background data pattern. We build a relationship between the memory cell＇s static noise margin （SNM） and the background data, and study the influence of irradiation on the probability density function of ASNM, which is the difference between two data sides＇ SNMs, to discuss the reason for pattern imprinting. Finally, we demonstrate that, for micron and deep sub-micron devices, the mechanism of pattern imprinting is the bias-dependent threshold shift of the transistor, but for a deep sub-micron device the shift results from charge trapping in the shallow trench isolation （STI） oxide rather than from the gate oxide of the micron-device.

Direct measurement and analysis of total ionizing dose effect on 130 nm PD SOI SRAM cell static noise margin

In this work, the total ionizing dose（TID） effect on 130 nm partially depleted（PD） silicon-on-insulator（SOI） static random access memory（SRAM） cell stability is measured. The SRAM cell test structure allowing direct measurement of the static noise margin（SNM） is specifically designed and irradiated by gamma-ray. Both data sides＇ SNM of 130 nm PD SOI SRAM cell are decreased by TID, which is different from the conclusion obtained in old generation devices that one data side＇s SNM is decreased and the other data side＇s SNM is increased. Moreover, measurement of SNM under different supply voltages（Vdd） reveals that SNM is more sensitive to TID under lower Vdd. The impact of TID on SNM under data retention Vddshould be tested, because Vddof SRAM cell under data retention mode is lower than normal Vdd.The mechanism under the above results is analyzed by measurement of I–V characteristics of SRAM cell transistors.

Heavy ion energy influence on multiple-cell upsets in small sensitive volumes:from standard to high energies

The 28 nm process has a high cost-performance ratio and has gradually become the standard for the field of radiation-hardened devices.However,owing to the minimum physical gate length of only 35 nm,the physical area of a standard 6T SRAM unit is approximately 0.16μm^(2),resulting in a significant enhancement of multi-cell charge-sharing effects.Multiple-cell upsets(MCUs)have become the primary physical mechanism behind single-event upsets(SEUs)in advanced nanometer node devices.The range of ionization track effects increases with higher ion energies,and spacecraft in orbit primarily experience SEUs caused by high-energy ions.However,ground accelerator experiments have mainly obtained low-energy ion irradiation data.Therefore,the impact of ion energy on the SEU cross section,charge collection mechanisms,and MCU patterns and quantities in advanced nanometer devices remains unclear.In this study,based on the experimental platform of the Heavy Ion Research Facility in Lanzhou,low-and high-energy heavy-ion beams were used to study the SEUs of 28 nm SRAM devices.The influence of ion energy on the charge collection processes of small-sensitive-volume devices,MCU patterns,and upset cross sections was obtained,and the applicable range of the inverse cosine law was clarified.The findings of this study are an important guide for the accurate evaluation of SEUs in advanced nanometer devices and for the development of radiation-hardening techniques.

Verification of SEU resistance in 65 nm high-performance SRAM with dual DICE interleaving and EDAC mitigation strategies

A dual double interlocked storage cell(DICE)interleaving layout static random-access memory(SRAM)is designed and manufactured based on 65 nm bulk complementary metal oxide semiconductor technology.The single event upset(SEU)cross sections of this memory are obtained via heavy ion irradiation with a linear energy transfer(LET)value ranging from 1.7 to 83.4 MeV/(mg/cm^(2)).Experimental results show that the upset threshold(LETth)of a 4 KB block is approximately 6 MeV/(mg/cm^(2)),which is much better than that of a standard unhardened SRAM with an identical technology node.A 1 KB block has a higher LETth of 25 MeV/(mg/cm^(2))owing to the use of the error detection and correction(EDAC)code.For a Ta ion irradiation test with the highest LET value(83.4 MeV/(mg/cm^(2))),the benefit of the EDAC code is reduced significantly because the multi-bit upset proportion in the SEU is increased remarkably.Compared with normal incident ions,the memory exhibits a higher SEU sensitivity in the tilt angle irradiation test.Moreover,the SEU cross section indicates a significant dependence on the data pattern.When comprehensively considering HSPICE simulation results and the sensitive area distributions of the DICE cell,it is shown that the data pattern dependence is primarily associated with the arrangement of sensitive transistor pairs in the layout.Finally,some suggestions are provided to further improve the radiation resistance of the memory.By implementing a particular design at the layout level,the SEU tolerance of the memory is improved significantly at a low area cost.Therefore,the designed 65 nm SRAM is suitable for electronic systems operating in serious radiation environments.

Direct measurement of an energy-dependent single-event-upset cross-section with time-of-flight method at CSNS

To predict the soft error rate for applications, it is essential to study the energy dependence of the single-event-upset(SEU) cross-section. In this work, we present a direct measurement of the SEU cross-section with the Back-n white neutron source at the China Spallation Neutron Source. The measured cross section is consistent with the soft error data from the manufacturer and the result suggests that the threshold energy of the SEU is about 0.5 Me V, which confirms the statement in Iwashita’s report that the threshold energy for neutron soft error is much below that of the(n, α) cross-section of silicon.In addition, an index of the effective neutron energy is suggested to characterize the similarity between a spallation neutron beam and the standard atmospheric neutron environment.

First principles simulation technique for characterizing single event effects

This paper develops a new simulation technique to characterize single event effects on semiconductor devices. The technique used to calculate the single event effects is developed according to the physical interaction mechanism of a single event effect. An application of the first principles simulation technique is performed to predict the ground-test single event upset effect on field-programmable gate arrays based on 0.25μm advanced complementary metal-oxidesemiconductor technology. The agreement between the single event upset cross section accessed from a broad-beam heavy ion experiment and simulation shows that the simulation technique could be used to characterize the single event effects induced by heavy ions on a semiconductor device.

Pattern dependence in synergistic effects of total dose on single-event upset hardness

The pattern dependence in synergistic effects was studied in a 0.18 μm static random access memory（SRAM） circuit.Experiments were performed under two SEU test environments：3 Me V protons and heavy ions.Measured results show different trends.In heavy ion SEU test,the degradation in the peripheral circuitry also existed because the measured SEU cross section decreased regardless of the patterns written to the SRAM array.TCAD simulation was performed.TIDinduced degradation in n MOSFETs mainly induced the imprint effect in the SRAM cell,which is consistent with the measured results under the proton environment,but cannot explain the phenomena observed under heavy ion environment.A possible explanation could be the contribution from the radiation-induced GIDL in pMOSFETs.

Small-Scale CMOS Pseudo SRAM Module Design

An approach to design small scale CMOS static random access memory (SRAM) is proposed. The design of address decoder, memory cell, and the layout are included. This approach adopts flip-flop array structure. The flip-flops are used as the storage cells and they are stacked to form the whole SRAM module. The word select bit is generated from the address decoder. And one word at a time is selected for reading or writing. The design of the memory core's layout is also discussed since it should be optimized to save area and also should be convenient for realization. It's a full-custom layout. The address decoder is composed of combinational logic circuit and its layout is also designed as a full-custom layout. With all these modules, the integral structure of the SRAM is carried out.

Analysis of functional failure mode of commercial deep sub-micron SRAM induced by total dose irradiation

Functional failure mode of commercial deep sub-micron static random access memory（SRAM） induced by total dose irradiation is experimentally analyzed and verified by circuit simulation. We extensively characterize the functional failure mode of the device by testing its electrical parameters and function with test patterns covering different functional failure modes. Experimental results reveal that the functional failure mode of the device is a temporary function interruption caused by peripheral circuits being sensitive to the standby current rising. By including radiation-induced threshold shift and off-state leakage current in memory cell transistors, we simulate the influence of radiation on the functionality of the memory cell. Simulation results reveal that the memory cell is tolerant to irradiation due to its high stability, which agrees with our experimental result.

Low complexity SEU mitigation technique for SRAM-based FPGAs

An internal single event upset（SEU）mitigation technique is proposed,which reads back the configuration frames from the static random access memory（SRAM）-based field programmable gate array（FPGA）through an internal port and compares them with those stored in the radiationhardened memory to detect and correct SEUs.Triple modular redundancy（TMR）,which triplicates the circuit of the technique and uses majority voters to isolate any single upset within it,is used to enhance the reliability.Performance analysis shows that the proposed technique can satisfy the requirement of ordinary aerospace missions with less power dissipation,size and weight.The fault injection experiment validates that the proposed technique is capable of correcting most errors to protect spaceborne facilities from SEUs.

Impacts of NBTI/PBTI on power gated SRAM

A signal probability and activity probability (SPAP) model was proposed firstly, to estimate the impacts of the negative bias temperature instability (NBTI) and positive bias temperature instability (PBTI) on power gated static random access memory (SRAM). The experiment results show that PBTI has significant influence on the read and write operations of SRAM with power gating, and it deteriorates the NBTI effects and results in a up to 39.38% static noise margin reduction and a 35.7% write margin degradation together with NBTI after 106 s working time. Then, a circuit level simulation was used to verify the assumption of the SPAP model, and finally the statistic data of CPU2000 benchmarks show that the proposed model has a reduction of 3.85% for estimation of the SNM degradation after 106 s working time compared with previous work.

Design of power balance SRAM for DPA-resistance

A power balance static random-access memory（SRAM） for resistance to differential power analysis（DPA） is proposed. In the proposed design, the switch power consumption and short-circuit power consumption are balanced by discharging and pre-charging the key nodes of the output circuit and adding an additional shortcircuit current path. Thus, the power consumption is constant in every read cycle. As a result, the DPA-resistant ability of the SRAM is improved. In 65 nm CMOS technology, the power balance SRAM is fully custom designed with a layout area of 5863.6 μm^2.The post-simulation results show that the normalized energy deviation（NED） and normalized standard deviation（NSD） are 0.099% and 0.04%, respectively. Compared to existing power balance circuits, the power balance ability of the proposed SRAM has improved 53%.

Impact of temperature on single event upset measurement by heavy ions in SRAM devices

The temperature dependence of single event upset （SEU） measurement both in commercial bulk and silicon on insulator （SOI） static random access memories （SRAMs） has been investigated by experiment in the Heavy Ion Research Facility in Lanzhou （HIRFL）. For commercial bulk SRAM, the SEU cross section measured by 12C ions is very sensitive to the temperature. The temperature test of SEU in SOl SRAM was conducted by 209Bi and 12C ions, respectively, and the SEU cross sections display a remarkable growth with the elevated temperature for 12C ions but keep constant for 209Bi ions. The impact of temperature on SEU measurement was analyzed by Monte Carlo simulation. It is revealed that the SEU cross section is significantly affected by the temperature around the threshold linear energy transfer of SEU occurrence. As the SEU occurrence approaches saturation, the SEU cross section gradually exhibits less temperature dependency. Based on this result, the experimental data measured in HIRFL was analyzed, and then a reasonable method of predicting the on-orbit SEU rate was proposed.

Design of replica bit line control circuit to optimize power for SRAM

A design of a replica bit line control circuit to optimize power for SRAM is proposed. The proposed design overcomes the limitations of the traditional replica bit line control circuit, which cannot shut off the word line in time. In the novel design, the delay of word line enable and disable paths are balanced. Thus, the word line can be opened and shut off in time. Moreover, the chip select signal is decomposed, which prevents feedback oscillations caused by the replica bit line and the replica word line. As a result, the switch power caused by unnec- essary discharging of the bit line is reduced. A 2-kb SRAM is fully custom designed in an SMIC 65-nm CMOS process. The traditional replica bit line control circuit and the new replica bit line control circuit are used in the designed SRAM, and their performances are compared with each other. The experimental results show that at a supply voltage of 1.2 V, the switch power consumption of the memory array can be reduced by 53.7%.

Characterization of single-event multiple cell upsets in a custom SRAM in a 65 nm triple-well CMOS technology

In this paper, the characterization of single event multiple cell upsets(MCUs) in a custom SRAM is performed in a 65 nm triple-well CMOS technology, and O(linear energy transfer(LET) = 3.1 Me V cm2/mg), Ti(LET = 22.2 Me V cm2/mg) and Ge(LET = 37.4 Me V cm2/mg) particles are employed. The experimental results show that the percentage of MCU events in total upset events is 71.11%, 83.47% and 85.53% at O, Ti and Ge exposures. Moreover, due to the vertical well isolation layout, 100%(O), 100%(Ti) and 98.11%(Ge) MCU cluster just present at one or two adjacent columns, but there are still 4 cell upsets in one MCU cluster appearing on the same word wire. The characterization indicates that MCUs have become the main source of soft errors in SRAM, and even though combining the storage array interleaving distance(ID) scheme with the error detection and correction(EDAC) technique, the MCUs cannot be completely eliminated, new radiation hardened by design techniques still need to be further studied.