- Sales Tel：
- Xi an
- Technical support
FPGA Interfaces Speeding Up
Jon Masters, a chief ARM architect at Red Hat, developed an FPGA accelerator called Trilby, he described at last week’s ARM Tech Con. “Ninety percent of the effort in using an FPGA accelerator is in interfacing to it -- that’s crazy it should be 10% -- what I’d like to see is a standard for the industry,” he said.
Today’s FPGA accelerators typically require some programming in Verilog, but that’s unacceptable, said Masters. A researcher at Microsoft raised a similar compliant more in an August 2014 paperdescribing work using FPGA accelerators in Microsoft’s data centers.
Microsoft and China’s Baidu are both exploring use of FPGAs on the servers, sparking interest in the area. Intel accelerated interest with its $16 billion bid in June to acquire Altera, which has led the way in moving its FPGAs to OpenCL.
Masters said he has talked to all FPGA makers and others about starting an initiative to define a programming interface for FPGA accelerators, probably based on OpenCL. Such an interface should include standard drivers that support PCI Express virtualization and be available for download on system boot up, he said.
After Masters’ talk, an engineering manager from storage giant EMC expressed interest in such an initiative. FPGAs are great for a limited set of uses, he said.
This week, IBM and Xilinx announced a multi-year strategic collaboration to jointly develop FPGA accelerators for OpenPower systems that target data centers and network function virtualization for carriers. They are collaborating on a cloud-based service for FPGA acceleration and will work on enabling programming FPGAs in C, C++ and OpenCL.
IBM is already delivering with Nvidia on a cloud-based GPU accelerator service. Two companies announced Power8 systems using Nvidia accelerators, something IBM said could dramatically boost response from its Watson system.
Data center giants so far say FPGAs are less expensive and power hungry than GPUs for accelerators. Microsoft and Baidu are pioneering new applications for the FPGAs in areas ranging from Web search to networking.
Paramesh Gopi, chief executive of Applied Micro, is one of the skeptics, seeing FPGA apps only useful for “a tiny sliver of vertical applications,” he said at ARM Tech Con. “What drives this market is open source software [and] there is no tie in to anything like an FPGA for that to take off,” he added.
Gopi is understandably defensive on the topic. His company is trying to mount an assault with ARM-based servers on Intel’s data center dominance.
From Masters’ point of view, the Intel/Altera combination is a call to the ARM community to develop its own answer to whatever x86/FPGA interfaces Intel/Altera define. IBM and Xilinx are clearly hammering out their own answer.