U.S. patent application number 15/929618 was filed with the patent office on 2021-11-18 for optimal placement of data structures in a hybrid memory based inference computing platform.
This patent application is currently assigned to INTERNATIONAL BUSINESS MACHINES CORPORATION. The applicant listed for this patent is INTERNATIONAL BUSINESS MACHINES CORPORATION. Invention is credited to Arvind KUMAR, Carl RADENS, Ashish RANJAN.
Application Number | 20210357138 15/929618 |
Document ID | / |
Family ID | 1000005940971 |
Filed Date | 2021-11-18 |
United States Patent
Application |
20210357138 |
Kind Code |
A1 |
RANJAN; Ashish ; et
al. |
November 18, 2021 |
OPTIMAL PLACEMENT OF DATA STRUCTURES IN A HYBRID MEMORY BASED
INFERENCE COMPUTING PLATFORM
Abstract
In a deep neural network (DNN), weights are defined that
represent a strength of connections between different neurons of
the DNN and activations are defined that represent an output
produced by a neuron after passing through an activation function
of receiving an input and producing an output based on some
threshold value. The weight traffic associated with a hybrid memory
therefore is distinguished from the activation traffic to the
hybrid memory, and one or more data structures may be dynamically
allocated in the hybrid memory according to the weights and
activations of the or more data structures in the DNN. The hybrid
memory includes at least a first memory and a second memory that
differ according to write endurance attributes.
Inventors: |
RANJAN; Ashish; (White
Plains, NY) ; KUMAR; Arvind; (Chappaqua, NY) ;
RADENS; Carl; (LaGrangeville, NY) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
INTERNATIONAL BUSINESS MACHINES CORPORATION |
Armonk |
NY |
US |
|
|
Assignee: |
INTERNATIONAL BUSINESS MACHINES
CORPORATION
Armonk
NY
|
Family ID: |
1000005940971 |
Appl. No.: |
15/929618 |
Filed: |
May 13, 2020 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 3/04 20130101; G06F
3/0604 20130101; G06F 3/0644 20130101; G06F 3/068 20130101 |
International
Class: |
G06F 3/06 20060101
G06F003/06; G06N 3/04 20060101 G06N003/04 |
Claims
1. A method for optimized placement of data structures in memory in
a computing environment by a processor, comprising: distinguishing,
by a memory controller, between weights and activations of one or
more data structures in a deep neural network (DNN) using flags
attached to the one or more data structures, the flags having a
first value indicative of the weights and a second value indicative
of the activations; and dynamically allocating and routing the one
or more data structures in a hybrid memory according to the flags
indicative of the weights and activations of the one or more data
structures in the DNN, wherein the hybrid memory includes at least
a first memory and a second memory that differ according to one or
more write attributes.
2. The method of claim 1, further including partitioning the hybrid
memory for the weights and activations of the one or more data
structures, wherein the one or more write attributes include write
energy, write latency, write endurance, or a combination
thereof.
3. (canceled)
4. The method of claim 1, wherein the distinguishing further
includes distinguishing between the weights and the activations of
the one or more data structures according to characteristics of a
read operation, a write operation, or a combination thereof.
5. The method of claim 1, further including storing the weights in
the first memory of the hybrid memory.
6. The method of claim 1, further including storing the activations
in the second memory of the hybrid memory.
7. The method of claim 1, further including minimizing write
operations of the one or more data structures to the hybrid memory
based upon the distinguishing of the weights and activations.
8. A system for optimized placement of data structures in memory in
a computing environment, comprising: one or more computers with
executable instructions that when executed cause the system to:
distinguish, by a memory controller, between weights and
activations of one or more data structures in a deep neural network
(DNN) using flags attached to the one or more data structures, the
flags having a first value indicative of the weights and a second
value indicative of the activations; and dynamically allocate and
route the one or more data structures in a hybrid memory according
to the flags indicative of the weights and activations of the one
or more data structures in the DNN, wherein the hybrid memory
includes at least a first memory and a second memory that differ
according to one or more write attributes.
9. The system of claim 8, wherein the executable instructions that
when executed cause the system to partition the hybrid memory for
the weights and activations of the one or more data structures,
wherein the one or more write attributes include write energy,
write latency, write endurance, or a combination thereof.
10. (canceled)
11. The system of claim 8, wherein the distinguishing further
includes distinguishing between the weights and the activations of
the one or more data structures according to characteristics of a
read operation, a write operation, or a combination thereof.
12. The system of claim 8, wherein the executable instructions that
when executed cause the system to store the weights in the first
memory of the hybrid memory.
13. The system of claim 8, wherein the executable instructions that
when executed cause the system to store the activations in the
second memory of the hybrid memory.
14. The system of claim 8, wherein the executable instructions that
when executed cause the system to minimize write operations of the
one or more data structures to the hybrid memory based upon the
distinguishing of the weights and activations.
15. A computer program product for optimized placement of data
structures in memory by a processor in a computing environment, the
computer program product comprising a non-transitory
computer-readable storage medium having computer-readable program
code portions stored therein, the computer-readable program code
portions comprising: an executable portion that distinguishes, by a
memory controller, between weights and activations of one or more
data structures in a deep neural network (DNN) using flags attached
to the one or more data structures, the flags having a first value
indicative of the weights and a second value indicative of the
activations; and an executable portion that dynamically allocates
and routes the one or more data structures in a hybrid memory
according to the flags indicative of the weights and activations of
the one or more data structures in the DNN, wherein the hybrid
memory includes at least a first memory and a second memory that
differ according to one or more write attributes.
16. The computer program product of claim 15, further including an
executable portion that partitions the hybrid memory for the
weights and activations of the one or more data structures, wherein
the one or more write attributes include write energy, write
latency, write endurance, or a combination thereof.
17. (canceled)
18. The computer program product of claim 15, wherein the
distinguishing further includes distinguishing between the weights
and the activations of the one or more data structures according to
characteristics of a read operation, a write operation, or a
combination thereof.
19. The computer program product of claim 15, further including an
executable portion that: stores the weights in the first memory of
the hybrid memory; and stores the activations in the second memory
of the hybrid memory.
20. The computer program product of claim 15, further including an
executable portion that minimizes write operations of the one or
more data structures to the hybrid memory based upon the
distinguishing of the weights and activations.
Description
BACKGROUND OF THE INVENTION
Field of the Invention
[0001] The present invention relates in general to computing
systems, and more particularly to, various embodiments for optimal
placement of data structures in a hybrid memory based inference
engine/computing platform in a computing environment by a
processor.
Description of the Related Art
[0002] In today's society, consumers, business persons, educators,
and others use various computing network systems with increasing
frequency in a variety of settings. Computer systems may be found
in the workplace, at home, or at school. Computer systems may
include data storage systems, or disk storage systems, to process
and store data. In recent years, both software and hardware
technologies have experienced amazing advancement. With the new
technology, more and more functions are added, and greater
convenience is provided for use with these computing systems.
SUMMARY OF THE INVENTION
[0003] Various embodiments for optimal placement of data structures
in a hybrid memory based inference engine/computing platform using
one or more processors in a computing environment are provided. In
one embodiment, by way of example only, a method for providing
optimal placement of data structures in memory in a computing
environment, again by a processor, is provided. One or more data
structures may be dynamically allocated in a hybrid memory
according to weights and activations of the one or more data
structures in a deep neural network ("DNN"). The hybrid memory
includes at least a first memory and a second memory that differ
according to write attributes. The first memory may be a
non-volatile memory ("NVM") based memory and the second memory may
be a CMOS-based memory.
BRIEF DESCRIPTION OF THE DRAWINGS
[0004] In order that the advantages of the invention will be
readily understood, a more particular description of the invention
briefly described above will be rendered by reference to specific
embodiments that are illustrated in the appended drawings.
Understanding that these drawings depict only typical embodiments
of the invention and are not therefore to be considered to be
limiting of its scope, the invention will be described and
explained with additional specificity and detail through the use of
the accompanying drawings, in which:
[0005] FIG. 1 is a block diagram depicting an exemplary computing
node according to an embodiment of the present invention;
[0006] FIG. 2 is an additional block diagram depicting an exemplary
cloud computing environment according to an embodiment of the
present invention;
[0007] FIG. 3 is an additional block diagram depicting abstraction
model layers according to an embodiment of the present
invention;
[0008] FIGS. 4A-4B are additional block diagrams depicting a deep
neural network architectures in which aspects of the present
invention may be realized;
[0009] FIG. 5 is an additional block depicting optimal placement of
data structures in a hybrid memory based inference engine/computing
platform in which aspects of the present invention may be
realized;
[0010] FIG. 6 is an additional block diagram depicting an exemplary
functional relationship and operations between various aspects of
the present invention; and
[0011] FIG. 7 is a flowchart diagram depicting an additional
exemplary method for optimally placing data structures in a hybrid
memory in a computing environment by a processor, again in which
aspects of the present invention may be realized.
DETAILED DESCRIPTION OF THE DRAWINGS
[0012] The present invention relates generally to the field of
artificial intelligence ("AI") such as, for example, machine
learning and/or deep learning. "Machine learning" is used to
broadly describe a primary function of electronic systems that
learn from data. In machine learning and cognitive science, neural
networks are a family of statistical learning models inspired by
the biological neural networks of animals, and in particular the
brain. Neural networks can be used to estimate or approximate
systems and functions that depend on a large number of inputs and
are generally unknown.
[0013] Neural networks use a class of algorithms based on a concept
of inter-connected "neurons." In a typical neural network, neurons
have a given activation function that operates on the inputs. By
determining proper connection weights (a process also referred to
as "training"), a neural network achieves efficient recognition of
a desired patterns, such as, for example, images and characters.
Oftentimes, these neurons are grouped into "layers" in order to
make connections between groups more obvious and to each
computation of values.
[0014] Deep-learning refers to a class of machine learning
algorithms that are based on the learning of multiple levels of
features or representations of a set of data. Current deep-learning
methods include using a cascade of multiple layers of nonlinear
processing units for feature extraction and transformation. Feature
extraction refers to the process of receiving an initial set of
measured data and building derived values (or features) to
facilitate subsequent learning and generalization steps. In many
cases, higher level features are derived from lower level features
to generate a hierarchical representation of the measured data and
the derived features.
[0015] Moreover, deep learning algorithms are based on distributed
representations. Distributed representations operate under the
assumption that observed (or measured) data are the result of
interactions of one or more factors organized into one or more
layers. Conceptually, deep learning introduces an additional
assumption that the layers of factors that interact to provide the
measured data are representative of levels of abstraction or
composition. Under this assumption, multiple layers and layer sizes
correspond to different amounts of abstraction.
[0016] Overall, deep learning has led to major advances in
optimizing decision making and extracting business value from
unstructured data such as, for example, images, audio, videos and
texts. However, while deep learning achieves increased computing
performance on a wide variety of data types and tasks, some
challenges, data structure placement and memory allocations issues
of Deep Neural Networks (DNNs) exists.
[0017] For example, a DNN places significant demand on the memory
subsystems with growing model sizes and datasets. A growing
fraction of processer chip area and energy in DNN accelerators is
spent on memories, which face challenges with technology scaling.
Consequently, emerging non-volatile memory (NVMs) technologies that
can improve the efficiency of memory systems while overcoming the
challenges of complementary metal-oxide semiconductor ("CMOS")
memories, are of significant interest. NVMs offer several unique
attributes such as, for example, near-zero leakage, and high
density that are beneficial for memory design.
[0018] However, many different challenges exist with NVMs. Some of
them include, for example, the NVM is 1) limited by the high energy
requirements, 2) high latency incurred during write operations, 3)
limited endurance, or 4) limited number of write cycles.
Accordingly, a need exists for a novel solution that can exploit
the strengths of NVMs while overcoming their weaknesses, thereby
achieving energy-efficient memory systems in DNN accelerators.
[0019] Thus, to overcome this challenge, the present invention
provides a novel solution by distinguishing the memory traffic
pertaining to model data structures (e.g., weights and biases) of
DNN from the memory traffic resulting from activation data
structures for each DNN layer during inference. That is, the
present invention may distinguish between weight memory traffic
with activation memory traffic. Said differently, the present
invention may distinguish between weight memory data structures
with activation memory data structures.
[0020] In an additional aspect, the present invention may
dynamically allocate one or more data structures in a hybrid memory
according to weights and activations of one or more data structures
in a deep neural network ("DNN"). The hybrid memory includes at
least a first memory and a second memory that differ according to
write attributes, characteristics, and/or properties such as, for
example, differences in write energy, write latency, and/or write
endurance. In one aspect, the non-volatile memory ("NVM") based
memory may be the first memory and the CMOS-based memory may be the
second memory. In one aspect, the present invention may store
(e.g., reads) the model parameters to (from) a first memory (e.g.,
a first NVM based memory), and uses the second memory (e.g., a
complementary metal-oxide semiconductor ("CMOS")-based memory) such
as, for example, a static random access memory ("SRAM") to store
the output activations. This memory partitioning achieves
significant energy efficiency by exploiting the varying intensity
of read/write operations in each DNN data structure.
[0021] Thus, the present invention may mitigate the performance (or
energy) penalty due to disparity in read and write latency (or
energy) and may also improve endurance (i.e., extend the lifetime
of the memory). In one aspect, as used herein, "endurance" may be
referred to by how many times a write operation may write to the
memory. Therefore, the present invention may extend to NVMs (or any
other memory) that may have a disparity in writes/read performance
or may have a disparity in endurance.
[0022] In an additional aspect, the present invention provides for
optimal placement of data structures in memory. One or more data
structures may be dynamically allocated in one or more discrete
and/or separate memories that may include 1) a non-volatile memory
("NVM") based memory and CMOS-based memory, 2) a hybrid memory that
includes NVM and CMOS memory (e.g., a hybrid memory that includes
the NVM and the CMOS memory, which are not distinct and separated),
and/or 3) one or more memory types that differ in various
attributes (e.g., write attributes) such as, for example, write
energy, write latency, write endurance, or a combination thereof.
In one aspect, the first memory (e.g., NVM) may include, but not
limited to, memristor, magnetoresistive RAM ("MRAM"), resistive RAM
("RRAM"), conductive bridging RAM ("CBRAM"), or phase change memory
("PCM") elements. The second memory (e.g., CMOS memory) may include
SRAM or other type of memory that have reduced the write energy,
write latency, write endurance, or a combination thereof.
[0023] Said differently, the hybrid memory may be comprised of an
NVM and a CMOS that are not distinct and separated (e.g., a single
memory including both the NVM and the CMOS). Alternatively, the
hybrid memory may be comprised of an NVM and a CMOS that are
distinct and separated (e.g., multiple memories that include the
NVM and the CMOS).
[0024] For example, in one aspect, one or more data structures may
be dynamically allocated in one or more types of memory
configurations/combinations (e.g., the discrete and/or separate
memories that may include 1) a non-volatile memory ("NVM") based
memory and CMOS-based memory or the hybrid memory comprised of the
NVM and the CMOS) according to weights and activations of the or
more data structures in a deep neural network ("DNN"). That is, the
weights in a neural network represent the strength of the
connections between the different neurons. Similarly, activations
represent the output produced by a neuron after it goes through a
process called an "activation function". An activation function may
be a function that takes in some input and produces some output
value based on some threshold. In one aspect, the hybrid memory may
include both non-volatile memory ("NVM") based memory and
CMOS-based memory. In another aspect, the hybrid memory may include
different types of memories or memory combinations that each differ
according to the write endurance attributes of the specific
memory.
[0025] In another aspect, the present invention may leverage the
varying read/write characteristics of DNN data structures such as,
for example by distinguishing between activations from weights when
reading from/writing to memory. It should be noted that weights and
activations have different characteristics. The weights during DNN
inference are read intensive, while activations are both read and
write intensive. The number of writes to the one or more discrete
and/or separate memories may be minimized to achieve and increase
energy-efficiency. The present invention may also mitigate a
performance penalty.
[0026] For example, consider a scenario where there is only
non-volatile memory ("NVM")-based memory. In this case, both
weights and activations may be written/read from the NVM-based
memory. Writes in NVMs may incur significant latency (e.g., greater
than five times ("5.times.") more than CMOS-based SRAM). Thus,
using mechanisms of the illustrated embodiments, avoiding the
writes to NVMs for the dynamic (read/write-intensive) activations
and redirecting those writes to CMOS-based SRAM results in much
lower write energy, write latency, write endurance, or a
combination thereof minimizing the impact on overall system
performance.
[0027] Also, as used herein, a computing system may include large
scale computing called "cloud computing" in which resources may
interact and/or be accessed via a communications system, such as a
computer network. Resources may be software-rendered simulations
and/or emulations of computing devices, storage devices,
applications, and/or other computer-related devices and/or services
run on one or more computing devices, such as a server. For
example, a plurality of servers may communicate and/or share
information that may expand and/or contract across servers
depending on an amount of processing power, storage space, and/or
other computing resources needed to accomplish requested tasks. The
word "cloud" alludes to the cloud-shaped appearance of a diagram of
interconnectivity between computing devices, computer networks,
and/or other computer related devices that interact in such an
arrangement.
[0028] It should be noted that one or more computations or
calculations may be performed using various mathematical operations
or functions that may involve one or more mathematical operations
(e.g., solving differential equations or partial differential
equations analytically or computationally, using addition,
subtraction, division, multiplication, standard deviations, means,
averages, percentages, statistical modeling using statistical
distributions, by finding minimums, maximums or similar thresholds
for combined variables, etc.).
[0029] As used herein, by way of example only, "best,"
"appropriate," and/or "optimize" may be used herein interchangeable
and refer to and/or defined as "maximize," "minimize," or attain
one or more specific targets, objectives, goals, or intentions.
"Best," "appropriate," and/or "optimize" may also refer to
maximizing a benefit to a user (e.g., maximize a workflow
pipeline). "Best," "appropriate," and/or "optimize" may also refer
to making the most effective or functional use of a situation,
opportunity, or resource.
[0030] Additionally, "best," "appropriate," and/or "optimize" may
need not refer to a best solution or result but may refer to a
solution or result that "is good enough" for a particular
application, for example. In some implementations, an objective is
to suggest a "best" combination of sections/items/defined objects,
but there may be a variety of factors that may result in alternate
suggestion of a combination of sections/items/defined
objects/activities/applications yielding better results. Thus, some
changes to the variety of factors may result in a jump from one
minimum/maximum to another minimum/maximum. In either case,
resulting suggestions of a combination of sections/items/defined
objects/activities/applications of the optimal/best executor nodes
may be determined by whatever "function" may be considered "good
enough," "substantially optimal," and/or "sufficiently good."
Herein, the terms "best," "appropriate," and/or "optimize" may also
refer to such results based on minima (or maxima, depending on what
parameters are considered in the optimization problem) for
suggesting of a combination of optimal/best applications, programs,
and/or computing components.
[0031] In an additional aspect, the terms "optimize" and/or
"optimizing" may refer to an operation performed in order to
achieve an improved result such as, for example, reduced execution
costs or increased resource utilization, whether or not the optimum
result is actually achieved. Similarly, the term "optimize" may
refer to a component for performing such an improvement operation,
and the term "optimized" may be used to describe the result of such
an improvement operation.
[0032] Other examples of various aspects of the illustrated
embodiments, and corresponding benefits, will be described further
herein.
[0033] It is understood in advance that although this disclosure
includes a detailed description on cloud computing, implementation
of the teachings recited herein are not limited to a cloud
computing environment and/or computing systems associated with one
or more vehicles. Rather, embodiments of the present invention are
capable of being implemented in conjunction with any other type of
computing environment now known or later developed.
[0034] Cloud computing is a model of service delivery for enabling
convenient, on-demand network access to a shared pool of
configurable computing resources (e.g., networks, network
bandwidth, servers, processing, memory, storage, applications,
virtual machines, and services) that can be rapidly provisioned and
released with minimal management effort or interaction with a
provider of the service. This cloud model may include at least five
characteristics, at least three service models, and at least four
deployment models.
[0035] Characteristics are as follows:
[0036] On-demand self-service: a cloud consumer can unilaterally
provision computing capabilities, such as server time and network
storage, as needed automatically without requiring human
interaction with the service's provider.
[0037] Broad network access: capabilities are available over a
network and accessed through standard mechanisms that promote use
by heterogeneous thin or thick client platforms (e.g., mobile
phones, laptops, and PDAs).
[0038] Resource pooling: the provider's computing resources are
pooled to serve multiple consumers using a multi-tenant model, with
different physical and virtual resources dynamically assigned and
reassigned according to demand. There is a sense of location
independence in that the consumer generally has no control or
knowledge over the exact location of the provided resources but may
be able to specify location at a higher level of abstraction (e.g.,
country, state, or datacenter).
[0039] Rapid elasticity: capabilities can be rapidly and
elastically provisioned, in some cases automatically, to quickly
scale out and rapidly released to quickly scale in. To the
consumer, the capabilities available for provisioning often appear
to be unlimited and can be purchased in any quantity at any
time.
[0040] Measured service: cloud systems automatically control and
optimize resource use by leveraging a metering capability at some
level of abstraction appropriate to the type of service (e.g.,
storage, processing, bandwidth, and active user accounts). Resource
usage can be monitored, controlled, and reported providing
transparency for both the provider and consumer of the utilized
service.
[0041] Service Models are as follows:
[0042] Software as a Service (SaaS): the capability provided to the
consumer is to use the provider's applications running on a cloud
infrastructure. The applications are accessible from various client
devices through a thin client interface such as a web browser
(e.g., web-based e-mail). The consumer does not manage or control
the underlying cloud infrastructure including network, servers,
operating systems, storage, or even individual application
capabilities, with the possible exception of limited user-specific
application configuration settings.
[0043] Platform as a Service (PaaS): the capability provided to the
consumer is to deploy onto the cloud infrastructure
consumer-created or acquired applications created using programming
languages and tools supported by the provider. The consumer does
not manage or control the underlying cloud infrastructure including
networks, servers, operating systems, or storage, but has control
over the deployed applications and possibly application hosting
environment configurations.
[0044] Infrastructure as a Service (IaaS): the capability provided
to the consumer is to provision processing, storage, networks, and
other fundamental computing resources where the consumer is able to
deploy and run arbitrary software, which can include operating
systems and applications. The consumer does not manage or control
the underlying cloud infrastructure but has control over operating
systems, storage, deployed applications, and possibly limited
control of select networking components (e.g., host firewalls).
[0045] Deployment Models are as follows:
[0046] Private cloud: the cloud infrastructure is operated solely
for an organization. It may be managed by the organization or a
third party and may exist on-premises or off-premises.
[0047] Community cloud: the cloud infrastructure is shared by
several organizations and supports a specific community that has
shared concerns (e.g., mission, security requirements, policy, and
compliance considerations). It may be managed by the organizations
or a third party and may exist on-premises or off-premises.
[0048] Public cloud: the cloud infrastructure is made available to
the general public or a large industry group and is owned by an
organization selling cloud services.
[0049] Hybrid cloud: the cloud infrastructure is a composition of
two or more clouds (private, community, or public) that remain
unique entities but are bound together by standardized or
proprietary technology that enables data and application
portability (e.g., cloud bursting for load-balancing between
clouds).
[0050] A cloud computing environment is service oriented with a
focus on statelessness, low coupling, modularity, and semantic
interoperability. At the heart of cloud computing is an
infrastructure comprising a network of interconnected nodes.
[0051] Referring now to FIG. 1, a schematic of an example of a
cloud computing node is shown. Cloud computing node 10 is only one
example of a suitable cloud computing node and is not intended to
suggest any limitation as to the scope of use or functionality of
embodiments of the invention described herein. Regardless, cloud
computing node 10 is capable of being implemented and/or performing
any of the functionality set forth hereinabove.
[0052] In cloud computing node 10 there is a computer system/server
12, which is operational with numerous other general purpose or
special purpose computing system environments or configurations.
Examples of well-known computing systems, environments, and/or
configurations that may be suitable for use with computer
system/server 12 include, but are not limited to, personal computer
systems, server computer systems, thin clients, thick clients,
hand-held or laptop devices, multiprocessor systems,
microprocessor-based systems, set top boxes, programmable consumer
electronics, network PCs, minicomputer systems, mainframe computer
systems, and distributed cloud computing environments that include
any of the above systems or devices, and the like.
[0053] Computer system/server 12 may be described in the general
context of computer system-executable instructions, such as program
modules, being executed by a computer system. Generally, program
modules may include routines, programs, objects, components, logic,
data structures, and so on that perform particular tasks or
implement particular abstract data types. Computer system/server 12
may be practiced in distributed cloud computing environments where
tasks are performed by remote processing devices that are linked
through a communications network. In a distributed cloud computing
environment, program modules may be located in both local and
remote computer system storage media including memory storage
devices.
[0054] As shown in FIG. 1, computer system/server 12 in cloud
computing node 10 is shown in the form of a general-purpose
computing device. The components of computer system/server 12 may
include, but are not limited to, one or more processors or
processing units 16, a system memory 28, and a bus 18 that couples
various system components including system memory 28 to processor
16.
[0055] Bus 18 represents one or more of any of several types of bus
structures, including a memory bus or memory controller, a
peripheral bus, an accelerated graphics port, and a processor or
local bus using any of a variety of bus architectures. By way of
example, and not limitation, such architectures include Industry
Standard Architecture (ISA) bus, Micro Channel Architecture (MCA)
bus, Enhanced ISA (EISA) bus, Video Electronics Standards
Association (VESA) local bus, and Peripheral Component
Interconnects (PCI) bus.
[0056] Computer system/server 12 typically includes a variety of
computer system readable media. Such media may be any available
media that is accessible by computer system/server 12, and it
includes both volatile and non-volatile media, removable and
non-removable media.
[0057] System memory 28 can include computer system readable media
in the form of volatile memory, such as random-access memory (RAM)
30 and/or cache memory 32. Computer system/server 12 may further
include other removable/non-removable, volatile/non-volatile
computer system storage media. By way of example only, storage
system 34 can be provided for reading from and writing to a
non-removable, non-volatile magnetic media (not shown and typically
called a "hard drive"). Although not shown, a magnetic disk drive
for reading from and writing to a removable, non-volatile magnetic
disk (e.g., a "floppy disk"), and an optical disk drive for reading
from or writing to a removable, non-volatile optical disk such as a
CD-ROM, DVD-ROM or other optical media can be provided. In such
instances, each can be connected to bus 18 by one or more data
media interfaces. As will be further depicted and described below,
system memory 28 may include at least one program product having a
set (e.g., at least one) of program modules that are configured to
carry out the functions of embodiments of the invention.
[0058] Program/utility 40, having a set (at least one) of program
modules 42, may be stored in system memory 28 by way of example,
and not limitation, as well as an operating system, one or more
application programs, other program modules, and program data. Each
of the operating system, one or more application programs, other
program modules, and program data or some combination thereof, may
include an implementation of a networking environment. Program
modules 42 generally carry out the functions and/or methodologies
of embodiments of the invention as described herein.
[0059] Computer system/server 12 may also communicate with one or
more external devices 14 such as a keyboard, a pointing device, a
display 24, etc.; one or more devices that enable a user to
interact with computer system/server 12; and/or any devices (e.g.,
network card, modem, etc.) that enable computer system/server 12 to
communicate with one or more other computing devices. Such
communication can occur via Input/Output (I/O) interfaces 22. Still
yet, computer system/server 12 can communicate with one or more
networks such as a local area network (LAN), a general wide area
network (WAN), and/or a public network (e.g., the Internet) via
network adapter 20. As depicted, network adapter 20 communicates
with the other components of computer system/server 12 via bus 18.
It should be understood that although not shown, other hardware
and/or software components could be used in conjunction with
computer system/server 12. Examples, include, but are not limited
to: microcode, device drivers, redundant processing units, external
disk drive arrays, RAID systems, tape drives, and data archival
storage systems, etc.
[0060] Referring now to FIG. 2, illustrative cloud computing
environment 50 is depicted. As shown, cloud computing environment
50 comprises one or more cloud computing nodes 10 with which local
computing devices used by cloud consumers, such as, for example,
personal digital assistant (PDA) or cellular telephone 54A, desktop
computer 54B, laptop computer 54C, and/or automobile computer
system 54N may communicate. Nodes 10 may communicate with one
another. They may be grouped (not shown) physically or virtually,
in one or more networks, such as Private, Community, Public, or
Hybrid clouds as described hereinabove, or a combination thereof.
This allows cloud computing environment 50 to offer infrastructure,
platforms and/or software as services for which a cloud consumer
does not need to maintain resources on a local computing device. It
is understood that the types of computing devices 54A-N shown in
FIG. 2 are intended to be illustrative only and that computing
nodes 10 and cloud computing environment 50 can communicate with
any type of computerized device over any type of network and/or
network addressable connection (e.g., using a web browser).
[0061] Referring now to FIG. 3, a set of functional abstraction
layers provided by cloud computing environment 50 (FIG. 2) is
shown. It should be understood in advance that the components,
layers, and functions shown in FIG. 3 are intended to be
illustrative only and embodiments of the invention are not limited
thereto. As depicted, the following layers and corresponding
functions are provided:
[0062] Device layer 55 includes physical and/or virtual devices,
embedded with and/or standalone electronics, sensors, actuators,
and other objects to perform various tasks in a cloud computing
environment 50. Each of the devices in the device layer 55
incorporates networking capability to other functional abstraction
layers such that information obtained from the devices may be
provided thereto, and/or information from the other abstraction
layers may be provided to the devices. In one embodiment, the
various devices inclusive of the device layer 55 may incorporate a
network of entities collectively known as the "internet of things"
(IoT). Such a network of entities allows for intercommunication,
collection, and dissemination of data to accomplish a great variety
of purposes, as one of ordinary skill in the art will
appreciate.
[0063] Device layer 55 as shown includes sensor 52, actuator 53,
"learning" thermostat 56 with integrated processing, sensor, and
networking electronics, camera 57, controllable household
outlet/receptacle 58, and controllable electrical switch 59 as
shown. Other possible devices may include, but are not limited to
various additional sensor devices, networking devices, electronics
devices (such as a remote-control device), additional actuator
devices, so called "smart" appliances such as a refrigerator or
washer/dryer, and a wide variety of other possible interconnected
objects.
[0064] Hardware and software layer 60 include hardware and software
components. Examples of hardware components include: mainframes 61;
RISC (Reduced Instruction Set Computer) architecture-based servers
62; servers 63; blade servers 64; storage devices 65; and networks
and networking components 66. In some embodiments, software
components include network application server software 67 and
database software 68.
[0065] Virtualization layer 70 provides an abstraction layer from
which the following examples of virtual entities may be provided:
virtual servers 71; virtual storage 72; virtual networks 73,
including virtual private networks; virtual applications and
operating systems 74; and virtual clients 75.
[0066] In one example, management layer 80 may provide the
functions described below. Resource provisioning 81 provides
dynamic procurement of computing resources and other resources that
are utilized to perform tasks within the cloud computing
environment. Metering and Pricing 82 provides cost tracking as
resources are utilized within the cloud computing environment, and
billing or invoicing for consumption of these resources. In one
example, these resources may comprise application software
licenses. Security provides identity verification for cloud
consumers and tasks, as well as protection for data and other
resources. User portal 83 provides access to the cloud computing
environment for consumers and system administrators. Service level
management 84 provides cloud computing resource allocation and
management such that required service levels are met. Service Level
Agreement (SLA) planning and fulfillment 85 provides
pre-arrangement for, and procurement of, cloud computing resources
for which a future requirement is anticipated in accordance with an
SLA.
[0067] Workloads layer 90 provides examples of functionality for
which the cloud computing environment may be utilized. Examples of
workloads and functions which may be provided from this layer
include: mapping and navigation 91; software development and
lifecycle management 92; virtual classroom education delivery 93;
data analytics processing 94; transaction processing 95; and, in
the context of the illustrated embodiments of the present
invention, various workloads and functions 96 for optimized
placement of data structures in a hybrid memory. In addition,
workloads and functions 96 for optimized placement of data
structures in a hybrid memory may include such operations as data
analytics, data analysis, and as will be further described, memory
allocation functionality. One of ordinary skill in the art will
appreciate that the workloads and functions 96 for optimized
placement of data structures in a hybrid memory may also work in
conjunction with other portions of the various abstractions layers,
such as those in hardware and software 60, virtualization 70,
management 80, and other workloads 90 (such as data analytics
processing 94, for example) to accomplish the various purposes of
the illustrated embodiments of the present invention.
[0068] Turning now to FIG. 4A, a block diagram depicts an
architecture of a deep neural network ("DNN") 400 in which
illustrative embodiments may be implemented. In one aspect, one or
more of the components, modules, services, applications, and/or
functions described in FIGS. 1-3 may be used in FIGS. 4A-4B.
[0069] For example, as depicted in FIG. 4A, the DNN 400 may be
components that may be incorporated, controlled, accessed, managed,
and/or included in a computer system/server 12 of FIG. 1. In one
aspect, repetitive description of like elements, components,
modules, services, applications, and/or functions employed in other
embodiments described herein (e.g., FIGS. 1-3) is omitted for sake
of brevity.
[0070] The DNN 400 may include an on-chip memory 404 (e.g.,
NVM-CMOS hybrid memory) to capture all different types/kinds of
storage and a compute core 406. The DNN 400 may be trained using
one or more operations such as, for example, backpropagation
("BP"). During the training, a large number of data structures may
be generated that are associated with each layer (e.g.,
convolutional layers such as Conv1, Conv2_1, Conv2_2, . . . ,
Conv5_9, and FC1) such as, for example, weights and activations.
That is, the DNN 400 may have a layered architecture where each
layer (e.g., Conv1, Conv2_1, Conv2_2, . . . , Conv5_9, and FC1) may
include some number of nodes, and each node effectively conducts
some mathematical operation on some input to calculate an
output.
[0071] As depicted, the DNN 400 may include one or more layers ("N
layers") such as, for example, layers 402 (e.g., convolutional
layers such as Conv1, Conv2_1, Conv2_2, . . . , Conv5_9, and fully
connected "FC1"), where some are convolutional layers and other
layers may be fully connected).
[0072] In one aspect, the layers 402 depicts an example DNN layer
with multiple layers such as, for example, Conv1, Conv2_1, Conv2_2,
. . . , Conv5_9, and fully connected "FC1").
[0073] An on-chip memory 404 and a compute core 406 depict a
typical DNN accelerator with an engine that executes the layers
402. In one aspect, the on-chip memory 404 and a compute core 406
may include, by way of example only, a CMOS-based and/or NVM-based
memory/hybrid memory to capture all different types/kinds of
storage, that brings in the weights associated with the different
layers of the layers 402 (e.g., i.e. Conv1, Conv2_1, etc.) and
input images for the first layer. The engine of the on-chip memory
404 and a compute core 406 then operates each layer of the layers
402 sequentially using the compute core 406, producing outputs
(referred to as activations) that are stored in memory for a given
layer. To execute a subsequent or following layer, the engine then
uses those stored activations along with the weights for its layer
and the execution continues.
[0074] Additionally, during the process of using BP for training
the DNN 400, one or more compute passes may occur per layer such
as, for example, Conv1, Conv2_1, Conv2_2, . . . , Conv5_9, and FC1.
The compute passes may include a forward pass (e.g., forward
propagation "FP"), the BP, and a weight update pass. For example,
initially, during a FP the weights of each layer such as, for
example, Conv1, Conv2_1, Conv2_2, . . . , Conv5_9, and FC1 are used
to generate the activations (e.g., output activations). As time
progresses, at each time step, the FP operations are completed at
each of the layers such as, for example, Conv1, Conv2_1, Conv2_2, .
. . , Conv5_9, and FC1.
[0075] The input to any given node is a weighted sum of the outputs
of the previous layer (plus a bias term usually equal to one or
zero). As depicted (via the dashed lines), some outputs of some
layers are required as inputs to other layers such as, for example,
output of Conv1 is also used as input to Conv2_2 where the output
of Conv2_1 is also used as input to Conv2_2. It is these weights
that are learned by during training. To learn these parameters, the
output of a training run is compared with the value, and the error
is backpropagated through the network to update the weights.
[0076] It should be noted that current accelerators to execute DNNs
can only accommodate smaller models (e.g., ResNet-50) and are
limited by on-chip SRAM capacity constraints such as, for example,
power and silicon area budget. Thus, the present invention targets
larger models for inference such as, for example, a ResNet-101
(with approximately 39.8 million "M" weights). The data structures
during inference provided activation and weights and the minibatch
size may be equal to 1 (e.g., latency constrained inference). Also,
the activation and weight precision may be 8-bits.
[0077] It should be noted, by way of example only, that an
inference may be mean that a trained DNN model is used to infer or
predict a given input (e.g., classify the input as "X" such as, an
cat or a "Y" such as a dog) based on images shown to the trained
DNN model, or translate a text from a first language to a second
language based on the input text. The inference may be the
production phase wherein a model may be used to predict real world
data.
[0078] Thus, as depicted, the model size of the DNN 400 may be 40
megabytes ("MB") that produces a maximum ("max") activation storage
of layers 402 such as, for example, an activation storage of 13
megabytes "MB". It should be noted that a model size may be the
total capacity of the model, i.e. the weights (and biases), while
the max activation storage represents what is the maximum size of
activations that need to be stored in memory. Thus, by applying the
mechanisms of the illustrated embodiments a higher density of
on-chip memory 404 enables larger on-chip buffer capacity. However,
the current state of the art stores weights and activation on an
NVM-based on-chip buffer 404 (e.g., compute core 406). The
non-hybrid NVMs lead to increased write energy and latency.
[0079] It should be noted that the non-hybrid memory (e.g., NVM
only) is less efficient than the hybrid NVM (e.g., the NVM and SRAM
as depicted in FIG. 5). For example, data mapping (e.g., data
placement) of the data structures is agnostic to the
characteristics of the DNN data structures. For example, the
weights are static (e.g., data is written once and read several
times). The activations are dynamic (e.g., data structures are
read/written several times). Thus, without the deployment of the
hybrid memory (see FIG. 5), a DNN employing the non-hybrid memory
yields increased energy inefficiency because of high write energy
of activations. For example, the NVM write energy is less efficient
as compared to the hybrid memory of the illustrated embodiments
(e.g., the NVM/SRAM read energy) because the NVM write energy is
multiple times higher (e.g., 3-5 times higher) and the activations
are written to (and read from) the NVM/CMOS buffer frequently.
Additionally, the low leakage benefits of NVMs are outweighed by
high write energy and performance penalty with slower write
operations since the off-chip memory latency is overlapped with
compute latency (e.g., the compute latency is hidden) and there is
reduced opportunity for latency hiding with slower writes.
[0080] Turning now to FIG. 4B, diagraph 415 depicts DNN 410 may
include one or more layers ("N layers") such as, for example,
layers 1-5 (e.g., where some are convolutional layers and other
layers may be fully connected). The DNN 410 may be trained using
one or more operations such as, for example, backpropagation
("BP"). During the training, a large number of data structures may
be generated that are associated with each layer (e.g., layers 1-5)
such as, for example, weights ("WEI") 440, activations ("ACT") 420,
deactivations ("DACT") 430, and weight gradients (WGRAD" or "WG")
450.
[0081] The DNN 410 may include a sequence of convolutional, and
fully connected layers. DNN inference uses the forward propagation
(FP) which involves two data structures; weights ("WEI") and
activations ("ACT"). In one aspect, two example types of
data-structures may be ACT/WEI that are equal to approximately 2N
data-structures for N layers, wherein "N" is a positive integer or
a defined value. Also, one compute passes per layer be may a
forward propagation that may be equal to approximately N compute
passes for N layers.
[0082] During the process of using backpropagation ("BP") for
training the DNN 410, one or more compute passes may occur per
layer, for example, layer 1 ("L1), layer 2, ("L2"), layer 3,
("L3"), layer 4, ("L4"), and/or layer 5, ("L5") (or layer N,
("LN"), where "N" is a positive integer). The compute passes may
include a forward pass (e.g., forward propagation "FP"), a backward
pass (e.g., backpropagation "BP"), and a weight update pass (e.g.,
weight-gradient). By way of example only, various dependencies
(shown in FIG. 4B only for illustrative purposes and not be
construed as limiting in nature) occurring during each time step
are depicted for the weights ("WEI") 440, the activations ("ACT")
420, the deactivations ("DACT") 430, and the weight gradients
(WGRAD" or "WG") 450 while training the DNN 410 using the
backpropagation ("BP").
[0083] For example, initially, during a forward pass ("FP"), the
weights of each layer (WEI 440) are used to generate the
activations (ACT 420). As time progresses at each time step and the
FP operations are completed at each of the layers, a series of
alternating steps may occur between the weight upgrade steps (e.g.,
WGRAD or WG 450) and the backward passes (e.g., backpropagation),
which also include various dependencies between each other. For
example, the weight upgrade steps (e.g., WGRAD or WG 450) may
require one or more of the activations (e.g., the ACT 420)
generated from the forward pass FP, which then generate the one or
more of the deactivations (e.g., the DACT 430) may be generated
from the BP.
[0084] Turning now to FIG. 5, a block diagram depicts an optimized
DNN architecture 500 with a hybrid memory for optimized placement
of data structures in the hybrid memory using a non-volatile deep
neural network inference operation in which illustrative
embodiments may be implemented. In one aspect, one or more of the
components, modules, services, applications, and/or functions
described in FIGS. 1-4 may be used in FIG. 5. For example, the
optimized DNN architecture 500 may be components that may be
incorporated, controlled, accessed, managed, and/or included in a
computer system/server 12 of FIG. 1. In one aspect, repetitive
description of like elements, components, modules, services,
applications, and/or functions employed in other embodiments
described herein (e.g., FIGS. 1-4) is omitted for sake of
brevity.
[0085] As illustrated, the optimized DNN architecture 500 may
include one or more types of memories such as, for example, an
first memory 510 (e.g., NVM), and a second memory 520 such as, for
example, a CMOS-based memory (e.g., SRAM), a hybrid controller 540,
and a compute core 530, each of which may in association and/or
communication. In one aspect, the first memory 510 (e.g., NVM) and
the second memory 520 (e.g., SRAM) create a hybrid memory. In one
aspect, the hybrid memory may include an on-chip NVM based buffer
and also an on-chip CMOS-based buffer. In one aspect, the DNN 500
may partition the memory (e.g., the first memory 510 (e.g., NVM)
and the second memory 520 (e.g., SRAM)) forming the hybrid memory
for storage of model parameters and output activations.
[0086] It should be noted that the first memory 510 (e.g., NVM) and
the second memory 520 (e.g., SRAM) may form a hybrid memory. The
hybrid memory may be one single memory that includes both the first
memory 510 and the second memory 520, which may be located on the
same computing system. Alternatively, the hybrid memory may be two
or more memories that includes both the first memory 510 and the
second memory 520, which are discrete and separated and may be
located on different or remote computing systems.
[0087] In one aspect, the "0" above the first memory 510 (e.g.,
NVM) and the "1" below the second memory 520 (e.g., SRAM) may be or
represent the flags or indicators (representing whether they are
activations ("ACT") or weights ("W")). If the value is `0` it means
the value represents a weight data structure, whereas if it is `1`
it would mean the value represent an activation. This indicator may
assist the controller to identify weights from activations and
accordingly redirect/reroute to the right destination, i.e. NVM or
CMOS.
[0088] In operation each read ("RD") and each write ("WR") may be
controlled by the hybrid controller 540. That is, the hybrid
controller 540 may determine and distinguish between the weights
and activations of the data structures ("DS"). For example, the
hybrid controller 540 may, for each RD, store (reads) the model
parameters to (and from) an NVM-based memory or "NVM" 510. The
hybrid controller 540 may use the CMOS-based memory (e.g., second
memory 520 (e.g., SRAM)) to store output activations to the
CMOS-based memory (e.g., second memory 520 (e.g., SRAM)). In an
additional aspect, the RD and WR represents the read and write
operation and depicts the act of reading and writing to the
memories (e.g., NVMs will always be just read, while SRAM could be
read or written). That is, output activations ("ACT") are both read
and written, while weights ("W") are just read.
[0089] In one aspect, the hybrid memory (e.g., the first memory 510
and the second memory 520) may include memory types that form one
single memory such as, for example, a single, hybrid memory in a
single computing system. Alternatively, the hybrid memory (e.g.,
the first memory 510 and the second memory 520) may include one or
more memory types that are distinct and separated from each other
such as, for example, in different computing systems.
[0090] In one aspect, the hybrid memory may be comprised of
elements such as, for example, SRAM. The first memory 510 (e.g.,
NVM-based memory) include elements including, but not limited to,
magnetorestive random-access memory ("MRAM"), memristor, resistive
RAM ("RRAM"), conductive fridging ("CBRAM"), or phase change memory
(PCM) elements. Thus, the hybrid controller 540 may provide for
dynamic (programmable) allocation and partitioning of the hybrid
memory (e.g., the first memory 510 (e.g., NVM) and the second
memory 520 (e.g., SRAM) for model parameters and output
activations. In an additional aspect, one or more compute cores
such as, for example, compute core 530 may be provide for DNN with
dynamic (programmable) allocation and partitioning of the hybrid
memory for model parameters and output activations between one or
more compute cores such as, for example, compute core 530.
[0091] Thus, the DNN 500 with hybrid memory may leverage the
varying read/write characteristics of DNN data structures such as,
for example by distinguishing between activations ("ACT") from
weights ("W") when reading from/writing to memory. The number of
writes ("WRs") to a first and a second memory (e.g., a second
memory with higher write energy) may be minimized to achieve and
increase energy-efficiency.
[0092] Turning now to FIG. 6, a block diagram depicting exemplary
functional components of system 600 according to various mechanisms
of the illustrated embodiments is shown. FIG. 6 illustrates a
system 600 for optimally placing data structures in a hybrid memory
in a computing environment. In one aspect, one or more of the
components, modules, services, applications, and/or functions
described in FIGS. 1-5 may be used in FIG. 6. In one aspect,
repetitive description of like elements, components, modules,
services, applications, and/or functions employed in other
embodiments described herein (e.g., FIGS. 1-5) is omitted for sake
of brevity.
[0093] With the foregoing in mind, the module/component blocks of
system 600 may also be incorporated into various hardware and
software components of a system for securing data compression in
accordance with the present invention. Many of the functional
blocks of system 600 may execute as background processes on various
components, either in distributed computing components, or on the
user device, or elsewhere.
[0094] Computer system/server 12 of FIG. 1 is shown incorporating a
DNN inference engine service 610. The DNN inference engine service
610 may incorporate processing unit 16 ("processor") and memory 28
of FIG. 1 to perform various computational, data processing and
other functionality in accordance with various aspects of the
present invention. The DNN inference engine service 610 may be
provided by the computer system/server 12 of FIG. 1.
[0095] As one of ordinary skill in the art will appreciate, the
depiction of the various functional units in the DNN inference
engine service 610 is for purposes of illustration, as the
functional units may be located within the DNN inference engine
service 610 or elsewhere within and/or between distributed
computing components.
[0096] In one aspect, the computer system/server 12 and/or the DNN
inference engine service 610 may provide virtualized computing
services (i.e., virtualized computing, virtualized storage,
virtualized networking, etc.). More specifically, the DNN inference
engine service 610 may provide, and/or be included in, a
virtualized computing, virtualized storage, virtualized networking
and other virtualized services that are executing on a hardware
substrate.
[0097] The DNN inference engine service 610 may include a
distinguishing component 640, a data placement component 650, a
partitioning component 660, and a memory component 670, each of
which may be in communication with each other.
[0098] In one aspect, distinguishing component 640 may distinguish
between the weights and the activations of the one or more data
structures for placing the one or more data structures in the
hybrid memory. The distinguishing component 640 may also
distinguish between the weights and the activations of the one or
more data structures according to characteristics of a read
operation, a write operation, or a combination thereof.
[0099] The data placement component 650 may dynamically allocate
one or more data structures in the memory component (e.g., a hybrid
memory) according to weights and activations of one or more data
structures in a deep neural network ("DNN"). In one aspect, the
memory component 670 is a hybrid memory that includes both
non-volatile memory ("NVM") based memory and CMOS-based memory
(e.g., SRAM).
[0100] The partitioning component 660 may partition the memory
component (e.g., a hybrid memory) for the weights and activations
of one or more data structures. The data placement component 650
may store the weights in the NVM based memory of the hybrid memory.
The data placement component 650 may store the activations in the
CMOS based memory of the hybrid memory. In an additional aspect,
the data placement component 650 may minimize write operations of
the one or more data structures to the non-volatile memory based
upon distinguishing the weights and activations.
[0101] Turning now to FIG. 7, an additional method 700 for
optimally placing data structures in a hybrid memory in a computing
environment by a processor is depicted, in which various aspects of
the illustrated embodiments may be implemented. The functionality
700 may be implemented as a method executed as instructions on a
machine, where the instructions are included on at least one
computer readable medium or one non-transitory machine-readable
storage medium. The functionality 700 may start in block 702.
[0102] Model parameters and activations of data structures may be
distinguished for placing the data structures in a hybrid memory,
as in block 704. The hybrid memory may be partitioned for the model
parameters and activations of the one or more data structures, as
in block 706. The data structures may be dynamically allocated in
the hybrid memory according to the model parameters and activations
of the one or more data structures in a deep neural network
("DNN"), as in block 708. The hybrid memory includes at least a
first memory and a second memory that differ according to write
attributes (e.g., write energy, a write latency, a write endurance,
or a combination thereof). The functionality 700 may end, as in
block 710.
[0103] In one aspect, in conjunction with and/or as part of at
least one block of FIG. 7, the operations of method 700 may include
each of the following. The operations of method 700 may dynamically
allocate the one or more data structures in the hybrid memory
according to weights and activations of one or more data structures
in the DNN. The one or more write attributes may include a write
energy, a write latency, a write endurance, or a combination
thereof. The hybrid memory includes may include a non-volatile
memory ("NVM") based memory and CMOS-based memory. The operations
of method 700 may partition the hybrid memory for the weights and
activations of one or more data structures.
[0104] The operations of method 700 may distinguish between the
weights and the activations of the one or more data structures for
placing the one or more data structures in the hybrid memory. The
operations of method 700 distinguish between the weights and the
activations of the one or more data structures according to
characteristics of a read operation, a write operation, or a
combination thereof.
[0105] The operations of method 700 store the weights in first
memory (e.g., the NVM based memory) of the hybrid memory, and store
the activations in the second memory (e.g., the CMOS based memory)
of the hybrid memory. The operations of method 700 minimize write
operations of the one or more data structures to the hybrid memory
based upon distinguishing the weights and activations. Thus, the
hybrid memory may include a first memory and a second memory that
differ in attributes, properties, and/or characteristics such as,
for example, write endurance, write energy and write latency. In
one aspect, the first memory may be an NVM and the second memory
may be a CMOS.
[0106] The present invention may be a system, a method, and/or a
computer program product. The computer program product may include
a computer readable storage medium (or media) having computer
readable program instructions thereon for causing a processor to
carry out aspects of the present invention.
[0107] The computer readable storage medium can be a tangible
device that can retain and store instructions for use by an
instruction execution device. The computer readable storage medium
may be, for example, but is not limited to, an electronic storage
device, a magnetic storage device, an optical storage device, an
electromagnetic storage device, a semiconductor storage device, or
any suitable combination of the foregoing. A non-exhaustive list of
more specific examples of the computer readable storage medium
includes the following: a portable computer diskette, a hard disk,
a random access memory (RAM), a read-only memory (ROM), an erasable
programmable read-only memory (EPROM or Flash memory), a static
random access memory (SRAM), a portable compact disc read-only
memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a
floppy disk, a mechanically encoded device such as punch-cards or
raised structures in a groove having instructions recorded thereon,
and any suitable combination of the foregoing. A computer readable
storage medium, as used herein, is not to be construed as being
transitory signals per se, such as radio waves or other freely
propagating electromagnetic waves, electromagnetic waves
propagating through a waveguide or other transmission media (e.g.,
light pulses passing through a fiber-optic cable), or electrical
signals transmitted through a wire.
[0108] Computer readable program instructions described herein can
be downloaded to respective computing/processing devices from a
computer readable storage medium or to an external computer or
external storage device via a network, for example, the Internet, a
local area network, a wide area network and/or a wireless network.
The network may comprise copper transmission cables, optical
transmission fibers, wireless transmission, routers, firewalls,
switches, gateway computers and/or edge servers. A network adapter
card or network interface in each computing/processing device
receives computer readable program instructions from the network
and forwards the computer readable program instructions for storage
in a computer readable storage medium within the respective
computing/processing device.
[0109] Computer readable program instructions for carrying out
operations of the present invention may be assembler instructions,
instruction-set-architecture (ISA) instructions, machine
instructions, machine dependent instructions, microcode, firmware
instructions, state-setting data, or either source code or object
code written in any combination of one or more programming
languages, including an object oriented programming language such
as Smalltalk, C++ or the like, and conventional procedural
programming languages, such as the "C" programming language or
similar programming languages. The computer readable program
instructions may execute entirely on the user's computer, partly on
the user's computer, as a stand-alone software package, partly on
the user's computer and partly on a remote computer or entirely on
the remote computer or server. In the latter scenario, the remote
computer may be connected to the user's computer through any type
of network, including a local area network (LAN) or a wide area
network (WAN), or the connection may be made to an external
computer (for example, through the Internet using an Internet
Service Provider). In some embodiments, electronic circuitry
including, for example, programmable logic circuitry,
field-programmable gate arrays (FPGA), or programmable logic arrays
(PLA) may execute the computer readable program instructions by
utilizing state information of the computer readable program
instructions to personalize the electronic circuitry, in order to
perform aspects of the present invention.
[0110] Aspects of the present invention are described herein with
reference to flowchart illustrations and/or block diagrams of
methods, apparatus (systems), and computer program products
according to embodiments of the invention. It will be understood
that each block of the flowchart illustrations and/or block
diagrams, and combinations of blocks in the flowchart illustrations
and/or block diagrams, can be implemented by computer readable
program instructions.
[0111] These computer readable program instructions may be provided
to a processor of a general-purpose computer, special purpose
computer, or other programmable data processing apparatus to
produce a machine, such that the instructions, which execute via
the processor of the computer or other programmable data processing
apparatus, create means for implementing the functions/acts
specified in the flowcharts and/or block diagram block or blocks.
These computer readable program instructions may also be stored in
a computer readable storage medium that can direct a computer, a
programmable data processing apparatus, and/or other devices to
function in a particular manner, such that the computer readable
storage medium having instructions stored therein comprises an
article of manufacture including instructions which implement
aspects of the function/act specified in the flowcharts and/or
block diagram block or blocks.
[0112] The computer readable program instructions may also be
loaded onto a computer, other programmable data processing
apparatus, or other device to cause a series of operational steps
to be performed on the computer, other programmable apparatus or
other device to produce a computer implemented process, such that
the instructions which execute on the computer, other programmable
apparatus, or other device implement the functions/acts specified
in the flowcharts and/or block diagram block or blocks.
[0113] The flowcharts and block diagrams in the Figures illustrate
the architecture, functionality, and operation of possible
implementations of systems, methods, and computer program products
according to various embodiments of the present invention. In this
regard, each block in the flowcharts or block diagrams may
represent a module, segment, or portion of instructions, which
comprises one or more executable instructions for implementing the
specified logical function(s). In some alternative implementations,
the functions noted in the block may occur out of the order noted
in the figures. For example, two blocks shown in succession may, in
fact, be executed substantially concurrently, or the blocks may
sometimes be executed in the reverse order, depending upon the
functionality involved. It will also be noted that each block of
the block diagrams and/or flowchart illustrations, and combinations
of blocks in the block diagrams and/or flowchart illustrations, can
be implemented by special purpose hardware-based systems that
perform the specified functions or acts or carry out combinations
of special purpose hardware and computer instructions.
* * * * *