Graphics Reference
In-Depth Information
Table 11.1
Comparison of vision processor approaches
Architecture
Pros
Cons
GPU
Good floating-point performance
Specialized programming model,
complex memory system
MPSoC
High performance at low power
Can't mix and match pieces
FPGA
High performance, leverages special-
ized numerical designs
Requires hardware expertise
compute division can play a key role in the cost/accuracy/performance trade-off
analysis. The VFloat library [
28
] provides a library of configurable floating-point
units for reconfigurable computing fabrics. The modules parameters can be used by
the designer to vary the accuracy and implementation cost of the generated module.
Table
11.1
summarizes some of the key differences between these approaches.
GPUs have high performance, particularly for floating-point algorithms, but have
a non-standard programming model and complex memory systems. MPSoCs have
good performance/power but the system designer must choose among platforms that
may vary widely in the accelerators they offer. FPGAs provide high performance,
particularly for specialized numerical algorithms, but their design requires hardware
expertise.
Platformdesign also needs efficient and scalableways to integrate accelerators and
processors. Networks are required for high-performance physical transport that meets
the high bandwidth demands of computer vision. Firmware and software interfaces
are crucial to the development and portability of systems based on such platforms.
Gudis et al. [
11
] developed a service-oriented framework for integrating accelerators
into heterogeneous processors. Video devices are connected by a crossbar. Vision
accelerators shared the same memory space with the ARM host processor. A vision
service framework provides abstractions for the accelerators and managing their
communication. Farabet et al. [
9
] used a dataflow style for accelerators connected in
a 2-D mesh. It uses a smart DMA unit as a memory controller.
11.3 Low-Power Camera Nodes
Low-power, small form factor cameras are increasingly available. The camera mod-
ules for hobbyist platforms such as Raspberry Pi are examples of the impressive
combinations of optics, image sensors, computers, and networks that we can build.
Researchers are also developing self-powered image sensors that can operate from
scavenged energy. Low-power camera nodes have several characteristics that influ-
ence the vision algorithms deployed on them. They may provide relatively low reso-
lution. They will almost certainly operate at small apertures to avoid focusing; this in
turn limits their low-light capabilities. They may have relatively simple processors,
largely due to limitations on the amount of heat they will be allowed to dissipate. If
they are self-powered, they will not always be on.
