When it comes to HPC, the accelerator market is largely dominated by GPUs, which are known for floating point performance prowess with IDC figures pointing to encroaching competition from Xeon Phi. FPGAs, while still a third-place contender, might have a rather remarkable year ahead in terms of their reach into wider markets, as well as in the “three Ps” arena with increasingly attractive price, performance, and programmability.
With the arrival of “big data” in HPC, hyperscale, and general commercial environments FPGAs have found a comfortable home in a broad range of applications that can be very broadly defined as “search” based—in other words, from actual web search (as is the case with Microsoft’s adoption of FPGAs to power Bing) to search and discovery applications in bioinformatics, security, image recognition and beyond. Last week at Hot Chips, Asian web giant Baidu revealed how they’re using field programmable gate arrays from Xilinx to boost a number of their pattern and image recognition capabilities with great success, and of course, they still have a place in one of the markets that’s always been a sweet spot in financial services.
It’s important to note that at this point, it’s not always a clear either/or question when it comes to GPUs and FPGAs in a number of application areas in this wide “search” classification. There are areas in what can broadly be termed “deep learning” that can benefit from both accelerator options—and to say nothing of the new wave of DSP-powered applications that will emerge. Although GPUs have a hefty head start in expanded markets with advanced programming tools in CUDA and OpenCL, efforts underway with OpenPower, Intel (which recently soft-announced its own FPGA-tuned chips) and ARM in particular could promise a brighter, broader landscape for FPGAs. Others, including Altera and Xilinx and specialized designs from companies like Adapteva are also on the watchlist for big things in 2015.
While a great deal of the more recent adoption of FPGAs has been centered on hyperscale and commercial work, there are still hot areas in HPC that are set to benefit from the expanded ecosystem around the accelerators. According to Convey’s CEO, Bruce Toal, they’re seeing great adoption in both hyperscale and research—particularly in the life sciences. And this is not expected to slow, even if their status as a vendor offering high performance FPGA-driven systems via their time-honed hybrid core integrated FPGA technology may change as new integrated offerings hit the shelves, especially those from close partner Intel.
“There’s never been more activity around FPGAs than we’re seeing now,” Toal said. “We’ve seen Moore’s Law have the expected impact on size and price performance so that the latest 20 and 28-nanometer FPGAs are really delivering a price performance point that hasn’t been possible to date—allowing very small devices to be integrated in a much higher volume sense in computing, rather than just on a specialized basis.”
While their software, tooling, and integration of the current hybrid core technology stands on its own for now, Toal said he’s recognized for a long time that FPGAs were heading toward further integration—a step that he isn’t resisting. While he’s thrilled to see the uptick in growth in capability and options for FPGAs, it does raise questions about how the business model for a company like Convey, which is selling integrated FPGAs that don’t offer quite the same tight integration that Intel and IBM. However, Toal said he is confident that their expertise and ability to adapt to the changing market will continue to place them ahead of the game.
To that end, Convey joined the OpenPower Foundation ranks to add their hybrid-core acceleration to the Power mix. “Our current portfolio of accelerated solutions coupled with the POWER architecture and IBM’s Coherent Accelerator Processor Interface (CAPI) offer data center customers the ability to deploy high performing, cost efficient solutions,” he said.
CAPI is that key dedicated piece of IP on the Power8 that exploits the coherency of the device itself in a very clean simple manner, which means it’s possible interact with the main memory of the processor in a low latency environment versus making unconnected hops. Although they haven’t shipped any systems sporting these new capabilities, Toal said he recognizes the high value of getting away from doing this over PCIe with cache coherency layered on top. He also noted that when IBM pulls through with the promise to keep pushing the memory bandwidth of Power8, performance will be taken to an entirely new level for users in hyperscale, bioinformatics and beyond.
“X86 has been dominant in HPC but the Power architecture has always been recognized as a capable architecture for HPC and other areas,” Toal noted. The OpenPower initiative has been gaining steam with notable sign-ins from Google and companies like Tyan building motherboards to meet demand, particularly in the hyperscale market. As for their HPC roots, Toal said there will be a great deal of interest going forward, especially as users begin to realize the programmability and features that are found in both integrated and co-processor approaches to FPGAs.
When that day comes, the real decisions about which FPGA-based solutions to ride with will come down to sheer price performance—not to mention end results. He said that even before the OpenPower CAPI work that’s been done, his team’s work on an image sizing application they deployed with partner Dell is getting a 48x performance improvement using their integrated hybrid core FPGA solution. When it’s all meshed together and tied with a CAPI string, the FPGA boxes of the future will be quite the present for the big search-based application areas in particular.
No comments:
Post a Comment