Microsoft Advances High-End Machine Learning with FPGA Accelerator

Microsoft researchers have made significant progress in developing a convolutional neural network (CNN) accelerator using field programmable gate arrays (FPGAs). The CNN accelerator can process image classification, image recognition, and natural language processing tasks at high speeds, with a reported throughput/watt three times higher than previous FPGA designs. This breakthrough is part of Microsoft's Catapult project, launched in 2011, which aims to leverage FPGAs to boost datacenter performance and reduce power consumption.

Key Takeaways:

  • Microsoft researchers have developed a CNN accelerator that runs at three times the throughput/watt of previous FPGA designs based on FPGAs.
  • The CNN accelerator significantly outperforms designs that use graphics processors, achieving higher performance and better energy efficiency.
  • The Catapult project, launched in 2011, aims to leverage FPGAs to boost datacenter performance, reduce power consumption, and develop new datacenter capabilities.
  • Microsoft plans to take the Catapult accelerator to production on Bing later this year, demonstrating its potential to accelerate cloud-based services.
  • The researchers have used the Arria 10 FPGA, which features hardened support for floating-point operations and runs at more than 1 Tflop with improved energy efficiency.

Statistics:

  • The CNN accelerator achieved a throughput/watt three times higher than previous FPGA designs.
  • The Arria 10 FPGA achieves datacenter performance up to 40 GFLOPS/watt, a significant increase in energy efficiency.
  • Microsoft estimates a 70-percent throughput increase using comparable energy with the Arria 10 FPGA.
  • The Catapult project has already demonstrated its potential by accelerating Bing web search ranking algorithms, with an FPGA board placed in each server.

Sources:

  • Microsoft white paper, "Convolutional Neural Network (CNN) Accelerator on FPGA", released on February 23.
  • Microsoft blog post, "Cake, Catalyst, and Cadence", by Doug Burger.
  • Altera press release, "Altera's Arria 10 FPGA Delivers Exceptional Performance and Power Efficiency for Datacenter Applications".
  • 2015 Global Data Point. All Rights Reserved. Provided by SyndiGate Media Inc. (Syndigate.info).