Tuesday, November 6, 2012

Altera announces industry’s first FPGA support for OpenCL

HONG KONG: Altera Corp. announced the FPGA industry’s first Software Development Kit (SDK) for OpenCL (Open Computing Language) which combines the massively parallel architecture of an FPGA with the OpenCL parallel programming model.

The SDK allows system developers and programmers familiar with C to quickly and easily develop high-performance, power-efficient FPGA-based applications in a high-level language. The Altera SDK for OpenCL enables FPGAs to work in concert with the host processor to accelerate parallel computation, at a fraction of the power compared to hardware alternatives. Altera will demonstrate the performance and productivity benefits of OpenCL for FPGAs at SuperComputing 2012 in booth #430.

“The industry’s approach for boosting system performance has evolved over time from increasing frequency in single-core CPUs, to using multi-core CPUs, to using parallel processor arrays,” said Vince Hu, vice president of product and corporate marketing at Altera. “This evolution leads us to today’s modern FPGAs, which are fine-grained, massively parallel digital logic arrays architected to execute computations in parallel. Our SDK for OpenCL enables customers to easily adopt FPGAs and leverage the performance and power benefits the devices provide.”

Altera SDK for OpenCL design flow
OpenCL is an open, royalty-free standard for cross-platform, parallel programming of hardware accelerators, including CPUs, GPGPUs and FPGAs.  The Altera SDK for OpenCL offers a unified, high-level design flow for hardware and software development that automates the time-consuming tasks required in typical hardware-design language (HDL) flows.

The OpenCL tool flow automatically converts OpenCL kernel functions into custom FPGA hardware accelerators, adds interface IPs, builds interconnect logic and generates the FPGA programming file. The SDK includes libraries that link to OpenCL API calls within a host program running on the CPU. By automatically handling these steps, designers are able to focus their development efforts on defining and iterating their algorithms rather than designing hardware.

The portability of the OpenCL code enables users to migrate their designs to different FPGAs or SoC FPGAs as their application requirements evolve. With SoC FPGAs, the CPU host is embedded into the FPGA, providing a single-chip solution that delivers significantly higher bandwidth and lower latency between the CPU host and the FPGA compared to using two discrete devices.

Using FPGAs to extract maximum parallelism in heterogeneous platforms
The Altera SDK for OpenCL enables programmers to leverage the massively parallel, fine-grained architectures featured in FPGAs to accelerate parallel computation. Unlike CPUs and GPGPUs, where parallel threads are executed across an array of cores, FPGAs allow kernel functions to be transformed into dedicated, deeply pipelined hardware circuits that are multithreaded using the concept of pipeline parallelism.

Each of these pipelines can be replicated many times to provide even more parallelism by allowing multiple threads to execute in parallel. The result is an FPGA-based solution that can deliver >5X performance/Watt compared to alternative hardware implementations.

Altera is working with several board partners to deliver COTS board solutions to customers. Currently, boards from BittWare and Nallatech are designed to support Altera OpenCL. Additional third-party boards will be supported with future releases of the SDK.

Altera has performed a variety of benchmarks that show the productivity savings and the performance and power efficiency gained by using an OpenCL framework for FPGA development. Based on early benchmarks and working with customers in a variety of markets, the SDK shaved months off one customer’s development time for their video processing application and boosted performance by 9X versus a CPU in another customer’s financial application.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.