Microsoft Accelerates Data Center Search with Altera Arria 10 FPGAs

Microsoft is leveraging Altera Arria 10 field programmable gate arrays (FPGAs) to achieve high-performance-per-Watt data center acceleration based on convolutional neural network (CNN) algorithms. By utilizing the Arria 10 FPGA's precision hard floating point digital signal processing (DSP) blocks and OpenCL coding, Microsoft researchers have demonstrated up to 40 GFLOPS-per-Watt in performance, a significant leap forward in CNN performance and power efficiency. This achievement marks a significant milestone in Microsoft's efforts to address data center infrastructure challenges through the adoption of reprogrammable FPGAs.

Key Takeaways:

  • Microsoft is using Altera Arria 10 FPGAs to accelerate CNN-based data center search, achieving up to 40 GFLOPS-per-Watt in performance.
  • The Arria 10 FPGA's precision hard floating point DSP blocks and support for OpenCL coding enable efficient CNN processing and data transfer.
  • Microsoft is experiencing a more than 3X performance-to-power advantage compared to general-purpose graphics processing units (GPGPUs) for CNN platforms.
  • Microsoft researchers are leveraging the Arria 10 Developer Kit and engineering samples to advance cloud technologies through reprogrammable FPGA adoption.
  • Altera's Stratix V FPGAs are being used to accelerate search on Microsoft's Catapult board in servers deployed in the first Bing data center.

Statistics:

  • 40 GFLOPS-per-Watt: the industry-leading level of performance achieved by Microsoft researchers using Altera Arria 10 FPGAs.
  • 3X: the performance-to-power advantage of Arria 10 FPGAs compared to GPGPUs for CNN platforms.
  • 25 February: the date of the Linley Group Data Center Conference in San Jose, California, where Altera will present on FPGAs accelerating data center search.

Sources:

  • Altera Corporation
  • Microsoft Corporation
  • FPGA 2015 ACM/SIGDA
  • Linley Group Data Center Conference
  • Microsoft blog post: "A Leap Forward in CNN Performance and Power Efficiency" by Doug Burger (director, Client and Cloud Apps, Microsoft Research)