Broadwell chip boosts GPU performance for COTS SBCs

Story

September 14, 2015

Aaron Frank

Curtiss-Wright

Embedded defense applications have the option of Broadwell chip, but may find challenges in thermal design

While Intel’s new Broadwell processor was developed for broad commercial markets, it also offers big benefits to embedded defense applications. Tech-refresh programs that take advantage of this new 5th-generation Core i7 processor will see significant performance-per-watt improvements, while new designs can exploit an extremely fast multicore processor and powerful integrated graphics GPU combined in a single package.

Broadwell is a die shrink of the previous-generation Haswell architecture chip. Intel, following a “Tick-Tock” model, alternates its processor development by creating a new microarchitecture in one generation and then shrinking the die geometry in the following generation. This Tick-Tock cadence steps forward roughly every 18 months. For general-purpose processing, Broadwell delivers a modest 10 to 15 percent improvement over the previous generation Haswell, or a more significant 40 to 50 percent improvement over the 3rd-generation Ivy Bridge family. A much larger payoff is evident in the chip’s graphics and floating point processing, driven by Broadwell’s integrated Graphics Processing Unit (GPU).

GPUs are massively parallel compute engines designed for intense graphics and floating-point math operations. At a high level, they lend themselves to two broad classes of processing, both of which are critical for today’s embedded defense applications: Rendering images for display and accelerated floating-point math operations for digital signal processing.

Broadwell’s predecessors, Ivy Bridge and Haswell, featured integrated GPUs that delivered an impressive 141 and 320 GFLOPS, respectively. While Broadwell’s final performance specifications are not yet public, its more powerful integrated GPU delivers a 40 to 60 percent GFLOP improvement over Haswell, and an even higher improvement over Ivy Bridge. This represents an enormous leap compared to what was available from a single piece of silicon just a few years ago, and is comparable to the best-in-class discrete GPU chips of just a few years ago. This new level of integrated GPU performance means that applications can drive multiple displays and/or accelerate DSP algorithms with the same piece of silicon that also manages general-purpose processing tasks. For many applications, this improved performance enables system designers to save a system slot or eliminate the use of a mezzanine graphics card. It also results in significant overall power reductions.

Broadwell features a 14-nm architecture, shrunk down from Ivy Bridge and Haswell’s 22-nm process. This reduced transistor feature size typically results in decreased power consumption, so algorithms designed for previous generation Core i7 processors can run at lower power levels with Broadwell. However, to increase GPU capabilities, Intel has packed more functionality into Broadwell, resulting in an overall chip power dissipation nearly identical to that of the 22-nm chips. For military commercial off-the-shelf (COTS) designers, this leads to a significant new challenge in the area of thermal design. Since power is being dissipated across a much smaller piece of silicon, this heat concentration can present an extremely difficult cooling challenge. The technology that previously adequately cooled Ivy Bridge and Haswell devices in demanding aerospace and defense environments has shown to be insufficient for Broadwell chips. In response, new cooling technology has been developed that specifically addresses the challenges of Broadwell 14-nm chips. In addition, enhanced power-saving configurability options are now available that enable designers to turn off processing cores or lower clock speeds to save power based on their application’s requirements. Using this approach, Broadwell-based SBCs can operate at as low as 16 W (typical) power, yet can also run at full power in the most challenging thermal environments.

Ideal for tech-refresh applications

Broadwell-based SBCs, ideal for technology-refresh applications, are well-suited to existing applications needing to run with less power in a smaller platform and in cases where enhancements to applications demand more than the original processor can handle.

Curtiss-Wright recently announced a new family of Broadwell-based SBCs (Figure 1). The family includes 3U and 6U VPX designs and a 6U VME design. An example of a Broadwell-based VPX SBC is the 3U OpenVPX VPX3-1259 module. This small board features a high-speed PCI Express Gen3 interface to the backplane. The new Broadwell SBC is 100 percent pin-compatible with previous-generation Ivy Bridge and Haswell SBCs.

Software support includes multiple versions of Linux (Fedora, Red Hat Enterprise Linux, Lynx Software’s LynxOS, and Concurrent Red Hawk), Wind River VxWorks, and Microsoft Windows Embedded.

Figure 1: The family of Broadwell-based SBCs includes 3U and 6U VPX designs as well as a 6U VME design.

(Click graphic to zoom by 1.9x)