IEEE Consumer Electronics Magazine - April 2017 - 54

Output

Unpooling

De-Conv

Unpooling

De-Conv

Pool/Conv

Pool

Conv Layer

Pool

Conv Layer

Input

Fully Connected Layers
Figure 10. an autoencoder is the merged structure of the encoder and decoder in a model.

the use of a scalable link interface
(SLI) bridge, but generic drivers currently exist only for two-GPU SLI for
titan X Pascal model 1080
model 1070
model 1060
the 10 series. Nvidia also sells nonconsumer GPUs that are specialized
Cuda cores
3,584
2,560
2,048
1,280
for deep learning that do not have
Memory capacity
12 gddr5X
8 gB gddr5X
8 gB gddr5
up to 6 gB gddr5
this driver limitation.
Memory speed
10 gb/s
10 gb/s
8 gb/s
8 gb/s
When choosing a GPU setup for
deep learning, it is important to
choose a device that has enough
memory
to
fit
your
model
and
enough cores to efficiently solve
optimized primarily for problems that are best solved sequentialthe problem. One also needs to decide on a multi-GPU setup verly. This is ideally suited for many common algorithms and tasks,
sus a single-GPU setup. It is also important that enough system
but is not performant for tasks with inherent parallelism. GPUs
memory exists to easily copy the model between GPU random
are designed for highly parallel graphical processes that can noraccess memory (RAM) and system RAM with room to spare for
mally be reduced to a limited set of matrix or tensor operations.
the operating system and any running applications and services.
As GPUs became more and more powerful to cope with the
increasing sophistication of 3-D graphical models used in gaming and in augmented and virtual reality, they became the predeep Learning on a StiCK
ferred option for deep learning research. A typical GPU has
Many companies are deploying hardware customized for deep
hundreds or thousands of cores, and although each core is much
learning, often targeted at deployment in smart objects. There
slower than a typical CPU core, together they are able to train
are significant speed and power consumption improvements that
networks (especially deep neural networks) at the level of one or
can be realized in hardware implementations of deep learning
two orders of magnitude faster than with a CPU. This makes
operations. Movidius, for example, released the Fathom Neural
intuitive sense. Human neurons are significantly slower than
Compute Stick, which uses approximately 12.5-times less power
GPU or CPU cores, but our brains are able to perform many recthan an equivalent Nvidia GPU in a small USB-compatible
ognition and classification tasks faster and with more accuracy
device [17] (Figure 11). It is targeted for use in robots, drones,
than most computer models. This is because human brains have
and surveillance cameras and can drastically improve recognibillions of individual asynchronous neurons and highly parallel
tion accuracy. It is a good choice for off-the-shelf deep learning
analog electrochemical communication.
hardware when one is ready to deploy a model.

Table 1. Nvidia's Pascal GPU Series [15], [16].

gpu verSuS Cpu training BenCHMarKS

Cudnn/Cuda

For some tasks, a GPU can be on the order of a hundred-times
faster than a CPU while using less power and at a much lower
cost. Generally, any task that requires the same operation to be
performed thousands of times (as in matrix multiplication) will
greatly benefit from the use of a GPU [14].

Nvidia is the most popular consumer GPU manufacturer with
deep learning researchers. Although their products are targeted
primarily at gaming and 3-D graphics processing, they also
invest significant resources in deep learning research. Nvidia
provides CUDNN, a low-level library of deep learning primitives that run on top of CUDA [17].

gpu ModeLS and Set-upS
Nvidia's latest consumer series is the Pascal series, labeled with a
"10" prefix (Table 1). This series represents a significant
improvement in both performance and power consumption over
previous models. Multiple Nvidia GPUs can be combined with
54 IEEE Consumer Electronics Magazine

APRIL 2017

deveLoping and evaLuating
deep Learning ModeLS
When deploying or designing a deep neural network, it is
useful to try out several variations with many parameters.

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - April 2017