Heterogeneous Manycore IP Framework
Cyceera’s vision for the future of computing is one that is cognitive, reconfigurable, highly parallel and heterogeneous. Legacy architectures and their programming models are no match for emerging complex, highly parallel computing and software applications. Cyceera has been developing a range of parallel processing and machine learning technologies, which exploit the features of our patented heterogeneous manycore technology and overcomes many of the limitations of current sequential processing technologies.
Cyceera’s patented parallel programming model provides a wealth of novel features that bring a mix of processor programmability, FPGA style flexibility and all with the performance characteristics of an ASIC or System on Chip (SoC) devices. Our technology enables the creation of powerful parallel processing devices that address the constantly increasing demands of reduced size, reduced power dissipation and greater throughput while offering flexible configurable hardware and targeted performance.
Modular and scalable spatial – dataflow inspired architecture comprising arrays of addressable fine and coarse grained function blocks connected by a simple Network on Chip (NoC) allowing hardware resources to be shared by different threads. The architecture limits data movement by processing the data where it is generated – stored where possible. Other key features are:
- Implements various levels of parallelism, including instruction, data, logic, thread, task, storage and IO.
- Fast hardware self-synchronizing thread mechanism that reduces overheads and is transparent to programmers.
- Optimizes hardware resource utilization through autonomous asynchronous scheduling and Out Of Order Execution (OoOE).
- Control hardware is also fragmented and shared by different threads to optimize utilization and reduce silicon real estate. Program code size and hence memory requirements are reduced.
Machine Learning IP Cores
The type of computation required for Machine Learning is quite different from computation for classical deterministic computing. Cyceera’s novel Machine Learning technology, which is not based on neural networks, provides a completely new computing platform, designed from the ground up for machine intelligence. By developing custom IP cores for Machine Learning, Cyceera is able to provide the means for its customers to tackle new research and increase their potential to do so much more with Machine Learning powered applications.
Cyceera’s Autonomous Real Time Machine Learning (ARTML) technology has a high functional density, a small silicon footprint and consumes little power. Cyceera’s Machine Learning technology can deliver orders of magnitude better-optimized performance per watt for Machine Learning than current solutions. It can handle both spatial and temporal information and can be used with our patented cognitive processing architecture.
As the common denominator in all these systems, Cyceera’s scalable Machine Learning technology provides an attractive solution that will significantly enhance future product development and performance. Low power consumption and a small silicon footprint allow our ARTML technology to be integrated into all manner of gadgets, wearables and sensors enabling them to operate intelligently and independently.