A c c e l e r a t e y o u r d e v e l o p m e n t Real Embedded - - PowerPoint PPT Presentation
A c c e l e r a t e y o u r d e v e l o p m e n t Real Embedded - - PowerPoint PPT Presentation
A c c e l e r a t e y o u r d e v e l o p m e n t Real Embedded company; 150 employees 120+ embedded software developers 20+ FPGA designers 10+ board designers Founded in 1996, privately owned Dutch based company, with
- Real Embedded company; 150 employees
– 120+ embedded software developers – 20+ FPGA designers – 10+ board designers
- Founded in 1996, privately owned
- Dutch based company, with offices in Best and Delft
- 3 Business unit:
– Consultancy: the Netherlands – Project execution: Europe and North America – Product development and sales: World Wide
- Xilinx Premier Alliance partner
2
Rieny Rijnen Founder & CEO
October 8th, 2015
- Closer integration of processing platforms
– System-on-Chip (Zynq, MPSoC) – Processor technologies cannot keep-up with Moore’s law; FPGA technologies can
FPGA integration in PC (PCI-e, Ethernet)
- Hardly any solution without processor system
– System-level approach needed for FPGA design – System-level architectures are mostly software driven =>Software- and hardware-world have to fully integrate
CPU performance (quad core i7) FPGA performance (Kintex Ultrascale)
FPGAs
1997 1999 2001 2002 2004 2005 2006 2009 2011 2012 2014 5000 500 50 5 5000 500 50 5
CPUs
8.2 TeraMACs (~1.5 TFlops) 72 GigaFlops
- Fixed functionality
– Functionality loaded at power-up – Every functions is always pre-loaded
- Fast parallel execution
– Massive processing power – High data throughput – Optimized pipe-line infrastructure
- Implementation of FPGA functionality
– Place & route very time consuming – One letter change in code means complete recalculation – Timing closure gets tougher while filling the FPGA
- No default infrastructure in place
- Operating system
- Dynamic task management/treading
- Memory management
- Process synchronization
- Queueing
- High level programming language
- One code base
- Short compile time
- Use PR to create and manage multiple PR
regions in FPGA
– Enables “software-like” dynamic task switching – Shorter compile times
- Use HLS
– High-level programming languages like C and C++ and OpenCL; “Spill a gate and win a week” – One code base for multiple processing platforms
- Reduce BOM cost due to smaller FPGA
- Smaller footprint; less complex board
- System remains operational while
functionality changes
- Ability to check underlying hardware by
swapping function-block with test-block and visa versa
- Additional effort required for creating “real”
OS like behavior:
– Predefined infrastructure around PR blocks including connection to CPU system – Queueing on FPGA for data synchronization – Event handling for process synchronization – Memory management for data sharing
- Keep your PR block within a clock region
- Keep the number of interconnects with PR low
- Play around with numbers, sizes and locations of PR’s
- Make PR reproducible; create a fixed framework
- Decouple clock domains by using FIFO’s between your
PR blocks and infrastructure
- Have a fast backplane for data exchange
=> this helps you to get a good PR design
SW Node interface (file) SW Node interface (file) SW Node interface (file) SW Node interface (file)
SW Node (process)
SW Node interface (file)
SW Node (process)
SW Node interface (file)
HW Node (proxy)
DYPLOLINK
Application
DYPLO CONNECT
AXI, PCIe, Ethernet, …
DYPLO BACKPLANE
(3 – 100Gbit/sec, e.g. 1xHDMI ~ 5Gbit/sec)
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
- 64Kx32bit memory mapped
memory space
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
DYPLO RECONFIGURABLE NODE
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
- 64Kx32bit memory mapped
memory space
DYPLO I/O NODE
- AXI stream inputs (3-
6Gbit/sec/stream) --AXI stream outputs (3- 6Gbit/sec/stream)
DYPLOLINK
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
- 64Kx32bit memory mapped
memory space
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
DYPLO FIXED NODE
- AXI stream inputs (3-6 Gbit/sec)
- AXI stream outputs (3-6 Gbit/sec)
- 64Kx32bit memory mapped
memory space
Eindhovenseweg 32-C, 5683 KH BEST, The Netherlands P.O. Box 440, 5680 AK BEST, The Netherlands Phone: +31 499 336969 | Fax: +31 499 336970 www.TopicProducts.com | info@TopicProducts.com