Scaling API-first – The story of a global engineering organization
Ximea - the pc camera, 90 gflps smart camera
1. The PC Camera
A New Class of Smart Camera
(…Or How to put 90 Gflops of Processing to Good Use)
VISION 2011, Stuttgart, November 10
2. Let’s Start with ‘Why’
XIMEA thinks you should be free to
demand cutting-edge performance,
industrial robustness and true
hardware/software compatibility from
your next compact vision system
without paying a premium.
3. Where the Machine Vision
Market Is Today
Maturity
=
Empowerment
=
Inflection Point
4. So What’s Next In the Evolution of
Machine Vision Systems?
5. First, Ask Yourself:
• How optimal is traditional integration of
components?
• Don’t we have huge overhead on
protocols/stacks/Links/MACs/PHYs?
• Plethora of interfaces, components, sparse soft-
and hardware-compatibility matrices
???WHY???
6. This ….. Not This
The PC Camera
A fully-functional, high-performance
industrial PC inside the camera
8. Aspects of PC Camera
• Fully optimized data path from sensor to the application
– Zero CPU overhead on image data delivery
– True zero copy paradigm
– Lowest possible latency
• Potential for integrated PLC to achieve sub-microsecond
jitter
• Complexity of hard- and software interfaces handled by
PC Camera vendor
11. PC Cameras Based on x86
• Sony, Matrox, NI, Leutron, Tattile, XIMEA all offer PC
Cameras
• Wealth of existing frameworks and applications (usually
tied to vendor’s full image processing library)
• Well-known operating systems (Linux, Windows,
Full/Embedded)
• Well-known application development tools (C++, etc.)
• New algorithms are first developed on PC, not limited to
sub-set of algorithms chosen by smart camera vendor
12. Atom PC Cameras –
Pinnacle of Perfection?
• Raw CPU performance in the range of 3GFlops
• What if you want to connect more than one camera?
– Runtime license cost
– High-speed interfaces are limited
• Upgradeability of RAM and SSD
13. Computing Platforms
We are here
Single-thread Performance
Enabled by:
• Rich data Parallelism
• Power-efficient GPUs
Constrained by:
• Power Constrained by:
• Parallel SW availability • Programming Models
• Scalability
Constrained by:
• Power
• Complexity
Single-core era Multi-core era Heterogeneous
computing era
14. New Era:
Heterogeneous computing
• APU – Accelerated Processing Units
• Collocating of CPU and GPU on single die
– CPU is used for OS and other infrastructure tasks
– GPU is used for number crunching
• Disadvantage of shared memory become an advantage
providing zero copy framework
• GPU is fully programmable with OpenCL and Direct-
Compute
20. CURRERA-G:
What It Means to You
• PC Camera with high performance processor made for
vector calculations and logic with true zero-copy memory
access
• Full OS or Embedded OS
• OS Adds Software Flexibility While Improving Remote
Support
• Lower latency than PC Host systems
• More than 25 API’s to the most popular image processing
libraries on the market
• And one or two other benefits…
21. Heat Issues: ✓
• Dissipating >20W from compact enclosure is
challenging and requires active cooling
• Micro heat-pipes
• Solid state microblowers
• Use of external connections
22. Embedded PLC vs. Latency: ✓
• Runs fully autonomous and independent
from main CPU and its OS
• Less than 1µs jitter provides higher
determinism than any RTOS can deliver
• Senses opto-isolated part or position
detector inputs
• Receives results of image processing
algorithm
• Controls opto-isolated outputs and
programmable LED light controller
• Graphical programming requires no
previous experience
• Programmable watchdog functionality,
can also reboot main CPU and its OS
24. What’s Next?
• Hardware AMD, 2012
– New A-Series APUs: Trinity, 32nm, 2.2GHz-3.1GHz, 2 and 4
cores
– Includes Turbo CORE and AMD Power Gating
– DDR3-2133, Radeon HD 7000
• Intel response?
• Software
– OpenCL infiltrates image processing libraries
– Development of task and data parallel computational algorithms
• Full integration of computational architecture and operating
systems.