Adapteva this week released the company’s Epiphany-IV chip, a coprocessor that can sit alongside the main processor in a data center and dramatically accelerate its computational performance at a whopping 72 GFLOPS per watt.
This fourth generation of the chip technology is produced via GlobalFoundries 28-nm manufacturing process. Inside, the Epiphany-IV includes 64 independent high-performance RISC cores on a single chip, delivering over 50 GHz of achievable programmable performance.
Specifically, each chip runs at 800 MHz, or the equivalent of 51.2 GHz. In total, the Epiphany delivers 100 GFLOPS, 72 GFLOPS/Watt at 500 MHz, and a 77,912 CoreMark score. At idle, the Epiphany-IV consumes just 80 milliwatts, with a “worst-case” maximum power consumption of 2 watts.
Adapteva has positioned the chip as an “application accelerator.” Placed next to an x86 or ARM microprocessor, it can help accelerate computationally intensive tasks. As such, it’s a sort of a competitor to a GPU, which some server providers have added to their offerings in order to speed calculations to a rate much faster than that of a general-purpose chip.
“The chip can scale up exponentially—1000s of cores working together,” a spokeswoman for Adapteva wrote in an email. “Anywhere that acceleration is an issue and parallel computing might be beneficial the Epiphany chip would be an ideal co-processor. It has extremely high level performance and extremely low energy needs.”
Although the chip is being positioned for the enterprise market, Adapteva chief executive Andreas Olofsson, claimed that the chip would also “enable server level performance in mobile devices such as smart-phones and tablets.” That’s an interesting reversal of the current path of chip evolution, in which mobile chips are tested out and inserted into servers to save power.
So far, the company has been relatively quiet in terms of design wins. The Adapteva spokeswoman said that the chip, once it reached third generation, has primarily been implemented in the military space. Adapteva has publicly worked with Bittware, for example, to produce a floating-point accelerator daughtercard to improve computational performance—a boon to HPC applications.
“It’s been in testing in a number of other marketplaces. The 28nm chip is necessary for an array of industries—and sampling means the ability to be implemented by a lot of the folks that have been ‘kicking the tires,’” the spokeswoman wrote.
In conjunction with the announcement of the Epiphany-IV, Adapteva also announced the early access release of its OpenCL SDK for the Epiphany multicore architecture. The OpenCL implementation was completed in conjunction with Brown Deer Technology, a software developer working to improve open-source heterogeneous computing. The Epiphany OpenCL SDK is currently in beta release and available to early access partners, according to Adapteva; code examples and other support documents will be published to the Adapteva site next month.