Kevin Brew

Kevin Brew

Auto-generated profile, based on publicly available information

Kevin Brew

IBM AI Hardware Center 2 , Albany, New York 12222

Research Focus & Metrics

Academic Career Length

10Years (2017-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
10 August 2023 in Nature Electronics
Verified
Journal Article
Closed access
PDF file
MGManuel Le Gallo
Manuel Le Gallo
RKRiduan Khaddam-Aljameh
Riduan Khaddam-Aljameh
MSMilos Stanisavljevic
Milos Stanisavljevic
AVAthanasios Vasilopoulos
Athanasios Vasilopoulos
BKBenedikt Kersting
Benedikt Kersting
ASAbu Sebastian
Abu Sebastian

23 August 2023 in Nature
Verified
Journal Article
Open access
PDF file
SAS. Ambrogio
S. Ambrogio
PNP. Narayanan
P. Narayanan
AOA. Okazaki
A. Okazaki
AFA. Fasoli
A. Fasoli
CMCharles Mackin
Charles Mackin
GBG. W. Burr
G. W. Burr
Models of artificial intelligence (AI) that have billions of parameters can achieve high accuracy across a range of tasks1,2, but they exacerbate the poor energy efficiency of conventional general-purpose processors, such as graphics processing units or central processing units. Analog in-memory computing (analog-AI)3–7 can provide better energy efficiency by performing matrix–vector multiplications in parallel on ‘memory tiles’. However, analog-AI has yet to demonstrate software-equivalent (SWeq) accuracy on models that require many such tiles and efficient communication of neural-network activations between the tiles. Here we present an analog-AI chip that combines 35 million phase-change memory devices across 34 tiles, massively parallel inter-tile communication and analog, low-power peripheral circuitry that can achieve up to 12.4 tera-operations per second per watt (TOPS/W) chip-sustained performance. We demonstrate fully end-to-end SWeq accuracy for a small keyword-spotting network and near-SWeq accuracy on the much larger MLPerf8 recurrent neural-network transducer (RNNT), with more than 45 million weights mapped onto more than 140 million phase-change memory devices across five chips.

11 October 2023 in IEEE Transactions on Electron Devices
Verified
Journal Article
Open access
AVAthanasios Vasilopoulos
Athanasios Vasilopoulos
JBJulian Büchel
Julian Büchel
BKBenedikt Kersting
Benedikt Kersting
CLCorey Lammie
Corey Lammie
KBKevin Brew
Kevin Brew
ASAbu Sebastian
Abu Sebastian
Analog in-memory computing (AIMC) using memristive devices is considered a promising Non-von Neumann approach for deep learning (DL) inference tasks. However, inaccuracies in the programming of devices, that are attributed to conductance variations, pose a key challenge toward achieving sufficient compute precision for DL inference. Fortunately, conduction variations in memristive devices, such as phase-change memory (PCM) devices, exhibit a strong state dependence. This state dependence can be exploited in synaptic unit cells that comprise more than one memristive device, to encode positive or negative weights. In such multi-memristive unit cells, we propose a method that optimally maps the weights to the device conductance values, by maximizing the number of devices at the stable SET and RESET states. We demonstrate that this method reduces the matrix-vector multiplication (MVM) error and is more resilient to non-ideal device retention characteristics. With this approach, we increase the mean experimental inference accuracy of a network trained for MNIST classification by 0.71% on two PCM-based AIMC cores, and the hardware-realistic simulated top-1 accuracy of a network trained for ImageNet classification by 0.28%, while significantly reducing variability across multiple experiment instances.

3 May 2017 in Journal of Applied Physics
Verified
Journal Article
Closed access
PDF file
TGTalia Gershon
Talia Gershon
OGOki Gunawan
Oki Gunawan
TGTayfun Gokmen
Tayfun Gokmen
KBKevin W. Brew
Kevin W. Brew
SSSaurabh Singh
Saurabh Singh
RHRichard Haight
Richard Haight
Recently, a new type of photovoltaic device based on an n-type Ag2ZnSnSe4 (AZTSe) absorber was demonstrated with an efficiency of over 5%. This work examines in detail several critical loss mechanisms in FTO/AZTSe/MoO3/ITO Schottky barrier devices. It is shown that the Schottky barrier height in the present devices under illumination is only ∼0.82 eV. With a barrier height of this magnitude, the device has an upper-bound of only 6.4% to the efficiency. Second, the AZTSe thin films used in the present devices exhibit exceptionally short minority carrier lifetimes (≤225 ps), as probed using two-photon spectroscopy. Third, hysteresis is observed in the J–V response of these devices. Pulsed electrical measurements reveal that reverse-bias voltage pulses create residual shunts in the device proportional to the pulse amplitude and duration. Reversible Ag-doping of the MoO3 buffer is proposed as an explanation of the observed shunting upon reverse bias. Therefore, improving the contact materials and the minority carrier lifetime is identified to be critically important for improving AZTSe-based photovoltaics. Finally, AZTSe is found to be photo-conductive with a Hall carrier density that increases by two orders of magnitude upon illumination. The origin of this effect is still unknown.

1 November 2023 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
Verified
Journal Article
Closed access
JBJulian Büchel
Julian Büchel
AVAthanasios Vasilopoulos
Athanasios Vasilopoulos
BKBenedikt Kersting
Benedikt Kersting
CLCorey Lammie
Corey Lammie
KBKevin Brew
Kevin Brew
ASAbu Sebastian
Abu Sebastian
Accurate programming of non-volatile memory (NVM) devices in analog in-memory computing (AIMC) cores is critical to achieve high matrix-vector multiplication (MVM) accuracy during deep learning inference workloads. In this paper, we propose a novel programming approach that directly minimizes the MVM error by performing stochastic gradient descent optimization with synthetic random input data. The MVM error is significantly reduced compared to the conventional unit-cell by unit-cell iterative programming. We demonstrate that the optimal hyperparameters in our method are agnostic to the weights being programmed, enabling large-scale deployment across multiple AIMC cores without further fine tuning. It also eliminates the need for high-resolution analog to digital converters (ADCs) to decipher the small unit-cell conductance during programming. We experimentally validate this approach by demonstrating an inference accuracy increase of 1.26% on ResNet-9. The experiments were performed using phase change memory (PCM)-based AIMC cores fabricated in 14nm CMOS technology.

1 December 2018
Verified
Conference Paper
Closed access
OGOleg Gluschenkov
Oleg Gluschenkov
HWHeng Wu
Heng Wu
KBKevin Brew
Kevin Brew
CNChengyu Niu
Chengyu Niu
LYLan Yu
Lan Yu
DGDechao Guo
Dechao Guo
We report on a significant pFET external resistance reduction (~40%) and corresponding 10% R ON decrease by nanosecond laser annealing of S/D structures applicable to advanced technology nodes. Selective melting of pFET S/D elements is responsible for this improvement. Process window boundaries are defined by channel and junction melting at the upper end and by S/D SiGe melting at the lower end. Short channel characteristics are not degraded within the identified process window. Contacted gate pitch (CPP) and fin number dependence of the process window is assessed.