Tuesday, May 04, 2021

VLSI Symposia: Why Sony Integrates MRAM on Sensor?

VLSI Symposia 2021 will be held in a fully virtual format due to COVID-19. While there are many imaging-related papers in the program, the most intriguing one comes from Sony on non-volatile MRAM integration onto an image sensor:
  • 3D Stacked CIS Compatible 40nm Embedded STT-MRAM for Buffer Memory,
    M. Oka*, Y. Namba*, Y. Sato*, H. Uchida*, T. Doi*, T. Tatsuno*, M. Nakazawa*, A. Tamura*, R. Haga*, M. Kuroda*, M. Hosomi*, K. Suemitsu**, E. Kariyada**, T. Suzuki**, H. Tanigawa**, M. Ueki**, M. Moritoki**, Y. Takegawa**, K. Bessho* and T. Umebayashi*,
    *Sony Semiconductor Solutions Corp. and
    **Sony Semiconductor Manufacturing Corp., Japan
    This paper presents the world's first demonstration of a 40nm embedded STT-MRAM for buffer memory, which is compatible with the 3D stacked CMOS image sensor (CIS) process. We optimized a CoFeB-based perpendicular magnetic tunnel junction (p-MTJ) to suppress the degradation of magnetic properties caused by the 3D stacked wafer process. With improved processes, we achieved high speed write operation below 40 ns under typical operation voltage conditions, endurance up to 1E+10 cycles and 1 s data retention required for a buffer memory. In addition, to broaden the application of embedded MRAM (eMRAM), we proposed a novel fusion technology that integrated embedded non-volatile memory (eNVM) and buffer memory type embedded MRAM in the same chip. We achieved a data retention of 1 s ~ >10 years with a sufficient write margin using the fusion technology.
Why does Sony need the MRAM on image sensor? My seven best guesses are:
  1. Column FPN calibration. Possibly, FPN can be reduced by another order of magnitude below its current level
  2. Dark current calibration. A few dark frames at different temperatures and exposure times can be stored on-chip, and subtracted, when needed.
  3. Per-pixel FPN calibration in global shutter pixels. For example, a storage node leakage can be measured at different temperatures and readout speeds and subtracted later. Or charge injection in voltage-domain GS pixels can be calibrated-out.
  4. Per-pixel PRNU and color crosstalk calibration to get a silky-smooth sky on photo
  5. PDAF pixel variations calibration, so that AF would be more accurate
  6. Temperature gradients and respective black level variations across the pixel array can be measured in different operating modes and temperatures and stored on-sensor
  7. Some kind of per-pixel calibration for ToF or event-driven sensors. Maybe store individual voltages for each APD in an array of APD pixels
Other image sensor-related papers are:
  • (Invited) A CMOS Image Sensor and an AI Accelerator for Realizing Edge-Computing-Based Surveillance Camera Systems,
    F. Morishita, N. Kato, S. Okubo, T. Toi, M. Hiraki, S. Otani, H. Abe, Y. Shinohara and H. Kondo, Renesas Electronics Corp., Japan
    This paper presents a CMOS image sensor and an AI accelerator to realize surveillance camera systems based on edge computing. For CMOS image sensors to be used for surveillance, it is desirable that they are highly sensitive even in low illuminance. We propose a new timing shift ADC used in CMOS image sensors for improving high sensitivity performance. Our proposed ADC improves non-linearity characteristics under low illuminance by 63%. Achieving power-efficient edge computing is a challenge for the systems to be used widely in the surveillance camera market. We demonstrate that our proposed AI accelerator performs inference processing for object recognition with 1 TOPS/W.
  • All-Directional Dual Pixel Auto Focus Technology in CMOS Image Sensors,
    E. S. Shim, Samsung Electronics Co., Ltd., Korea
    We developed a dual pixel with accurate and all-directional auto focus (AF) performance in CMOS image sensor (CIS). The optimized in-pixel deep trench isolation (DTI) provided accurate AF data and good image quality in the entire image area and over whole visible wavelength range. Furthermore, the horizontal-vertical (HV) dual pixel with the slanted in-pixel DTI enabled the acquisition of all-directional AF information by the conventional dual pixel readout method. These technologies were demonstrated in 1.4um dual pixel and will be applied to the further shrunken pixels.
  • Development of Advanced Inter-Color-Filter Grid on Sub-Micron-Pixel CMOS Image Sensor for Mobile Cameras with High Sensitivity and High Resolution,
    J. In-Sung, Y. Lee, H. Y. Park, J. U. Kim, D. Kang, T. Kim, M. Kim, K. Lee, M. Heo, I. Ro, J. Kim, I. Park, S. Kwon, K. Yoon, D. Park, C. Lee, E. Jo, M. Jeon, C. Park, K. R. Byun, C. K. Chang, J. S. Hur, K. Yoon, T. Jeon, J. Lee, J. Park, B. Kim, J. Ahn, H. Kim, C.-R. Moon and H.-S. Kim, Samsung Electronics Co., Ltd., Korea
    Sub-micron pixels have been widely adopted in recent CMOS image sensors to implement high resolution cameras in small form factors, i.e. slim mobile-phones. Even with shrinking pixels, customers demand higher image quality, and the pixel performance must remain comparable to that of the previous generations. Conventionally, to suppress the optical crosstalk between pixels, a metal grid has been used as an isolation structure between adjacent color filters. However, as the pixel size continues to shrink to the sub-micron regime, an optical loss increases because the focal spot size of the pixel's microlens does not downscale accordingly with the decreasing pixel size due to the diffraction limit: the light absorption inevitably occurs in the metal grid. For the first time, we have demonstrated a new lossless, dielectric-only grid scheme. The result shows 29 % increase in sensitivity and +1.2-dB enhancement in Y-SNR when compared to the previous hybrid metal-and-dielectric grid.
  • A 2.6 e-Rms Low-Random-Noise, 116.2 mW Low-Power 2-Mp Global Shutter CMOS Image Sensor with Pixel-Level ADC and In-Pixel Memory,
    M.-W. Seo, M. Chu, H.-Y. Jung, S. Kim, J. Song, J. Lee, S.-Y. Kim, J. Lee, S.-J. Byun, D. Bae, M. Kim, G.-D. Lee, H. Shim, C. Um, C. Kim, I.-G. Baek, D. Kwon, H. Kim, H. Choi, J. Go, J. Ahn, J.-k. Lee, C. Moon, K. Lee and H.-S. Kim, Samsung Electronics Co., Ltd., Korea
    This paper presents a low-random noise of 2.6 e-rms, a low-power of 116.2 mW at video rate, and a high-speed up to 960 fps 2-mega pixels global-shutter type CMOS image sensor (CIS) using an advanced DRAM technology. To achieve a high performance global-shutter CIS, we proposed a novel architecture for the digital pixel sensor which is a remarkable global shutter operation CIS with a pixel-wise ADC and an in-pixel digital memory. Each pixel has two small-pitch Cu-to-Cu interconnectors for the wafer-level stacking, and the pitch of each unit pixel is less than 5 um which is the world's smallest pixel embedding both pixel-level ADC and 22-bit memories.
  • A Photon-Counting 4Mpixel Stacked BSI Quanta Image Sensor with 0.3e- Read Noise and 100dB Single-Exposure Dynamic Range,
    J. Ma, D. Zhang, O. Elgendy and S. Masoodian, Gigajot Technology Inc., USA
    This paper reports a 4Mpixel, 3D-stacked backside illuminated Quanta Image Sensor (QIS) with 2.2um pixels that can operate simultaneously in photon-counting mode with deep sub-electron read noise (0.3e- rms) and linear integration mode with large full-well capacity (30k e-). A single-exposure dynamic range of 100dB is realized with this dual-mode readout under room temperature. This QIS device uses a cluster-parallel readout architecture to achieve up to 120fps frame rate at 550mW power consumption.
  • A 5.1ms Low-Latency Face Detection Imager with In-Memory Charge-Domain Computing of Machine-Learning Classifiers,
    H. Song*, S. Oh*, J. Salinas*, S.-Y. Park** and E. Yoon*, *Univ. of Michigan, USA and **Pusan National Univ., Korea
    We present a CMOS imager for low-latency face detection empowered by parallel imaging and computing of machinelearning (ML) classifiers. The energy-efficient parallel operation and multi-scale detection eliminate image capture delay and significantly alleviate backend computational loads. The proposed pixel architecture, composed of dynamic samplers in a global shutter (GS) pixel array, allows for energy-efficient in-memory charge-domain computing of feature extraction and classification. The illumination-invariant detection was realized by using log-Haar features. A prototype 240x240 imager achieved an on-chip face detection latency of 5.1ms with a 97.9% true positive rate and 2% false positive rate at 120fps. Moreover, a dynamic nature of in-memory computing allows an energy efficiency of 419pJ_pixel for feature extraction and classification, leading to the smallest latency-energy product of 3.66ms.nJ_pixel with digital backend processing.
  • A CMOS LiDAR Sensor with Pre-Post Weighted-Histogramming for Sunlight Immunity Over 105 klx and SPAD-Based Infinite Interference Canceling,
    S. Hyeongseok, Sungkyunkwan Univ., Korea
    This paper presents a CMOS LiDAR sensor with high background noise (BGN) immunity. The sensor has on-chip pre-post weighted histogramming to detect only time-correlated time-of-flight (TOF) out of BGN from both sunlight and exponentially increased dark noise while enhancing sensitivity through higher excess voltage (Vex) of SPADs. The sensor also employs a SPAD-based random number generator (SRNG) for canceling interference (IF) from an infinite number of LiDARs. The sensor shows 8.08 cm accuracy for the range of 32 m under high BGN (105 klx sunlight and 48.72 kcps dark-count rate with increased Vex).
  • Advanced Multi-NIR Spectral Image Sensor with Optimized Vision Sensing System and Its Impact on Innovative Applications,
    H. Sumi*, **, H. Takehara**, J. Ohta** and M. Ishikawa*, *The Univ. of Tokyo and **Nara Institute of Science and Technology, Japan
    Innovative applications with multiple near-infrared (multi-NIR) spectral CMOS image sensors (CIS) and camera systems have recently been developed. The multi-NIR filter is an indispensably key technology in practical of using the multi-NIR camera system in consumer camera. Advanced processing technology for multi-NIR signals has been developed using a Fabry-Perot structure. Three types of NIR wavelength filters are formed as a Bayer pattern with 2-x-2um2 pixel size on a 5-M pixel BSICIS. The thickness differences of the three types of bandpass filters are suppressed to less than 75 nm. To enable applications in surveillance, automobiles, and fundus cameras for health management, signal processing technology has also been developed that processes and mixes each signal of a multi-NIR signal with low-intensity visible light images. This provides good image SNR (Signal-to-Noise Ratio ) under low lighting conditions of 0.1 lux or less allowing changes of state to be easily identified.
  • Multiplex PCR CMOS Biochip for Detection of Upper Respiratory Pathogens including SARS-CoV-2,
    A. Manickam, K. A. Johnson, R. Singh, N. Wood, E. Ku, A. Cuppoletti, M. McDermott and A. Hassibi, InSilixa, Inc., USA
    A 1024-pixel CMOS biochip for multiplex polymerase chain reaction application is presented. Biosensing pixels include 137dB DDR photosensors and an integrated emission filter with OD ~ 6 to perform real-time fluorescence-based measurements while thermocycling the reaction chamber with heating and cooling rates of > ±10°C/s. The surface of the CMOS IC is biofunctionalized with DNA capturing probes. The biochip is integrated into a fluidic consumable enabling loading of extracted nucleic acid samples and the detection of upper respiratory pathogens, including SARS-CoV-2.
The Symposia also offers a short course "Image Sensor Technologies for Computer Vision Systems to Realize Smart Sensing" by A. Nose, Sony Semiconductor Solutions Corp.


  1. isn't it just needed for image storage for fast rolling shutter instead of DRAM?

    1. MRAM is non-volatile and they target 10 year retention time. Looks like an overkill for just a frame storage.

  2. Why this information is not stored on the HD of the PC which do the vendor specific raw conversion? Storing some dark images with exposure variation would be easy if there is a fresh memory card to format.

  3. Wow, this looks like a really nice session. Thanks for posting it.

    My guess about the MRAM is that it is a driver for future stacked CNN processing where they will need a lot of weight storage for the learned video processing.

  4. Very interesting - there are so many things that could be done with this, even the short retention time version.

    If the memory can undergo a read-modify-write cycle at the full framerate then all kinds of temporal filtering could be implemented, including e.g. change detection, allowing only those pixels with significant changes to be output from the device, or linear HDR modes where only a single frame needs to be read out rather than all of the multiple exposures (there are various ways in which that could be done...).

    Moving the high bandwidth activities onto the imager to save bandwidth at the system level could be a motivation?

    For the long retention variant, having the ability to store any sensor-specific per-pixel calibrations directly on the sensor is a very neat solution from the manufacturing perspective. Electro-optic calibration could be done during wafer sort (which already requires extensive EO test capabilities) and programmed to the device itself so it automatically 'travels' with the sensor chip, rather than having to be tracked and programmed to some system-level non-volatile storage later.

    If this technology is cheaply and reliably manufacturable and the power consumption is OK (should be) then I can see several ways it could help reduce system power / size / complexity / cost. Less exciting than CNN weight storage but probably good business for Sony...


All comments are moderated to avoid spam and personal attacks.