GTI GAINBOARD Inference Accelerator Server upgrades data center AI modeling at low power
December 27, 2018
The second-generation GAINBOARD 2803S delivers 1,084 TOPS of performance at 110 W, which GTI claims is the industry?s best performance-to-power ratio.
Gyrfalcon Technology Inc. (GTI) has released the GAINBOARD 2803S Inference Accelerator Server, an upgrade solution for data centers running large numbers of data models in parallel. The second-generation GAINBOARD 2803S delivers 1,084 TOPS of performance at 110 W, which GTI claims is the industry’s best performance-to-power ratio.
The GAINBOARD 2803S Inference Accelerator Server is based on dual Intel Xeon processors and four PCIe cards, each of which contains 16 GTI Lightspeeur 2803S chips and four Xilinx FPGAs. The PCIe accelerator cards can operate on multiple models in parallel, or in a ‘cascade’ mode that permits large and complex models to be spread across individual Lightspeeur chips. In cascade mode, workloads are spread across successive chips without the intervention of the host processor.
The Lightspeeur chips provide 16.8 TOPS at 0.7W with 2 millisecond responsiveness. More information on the Lightspeeur chip is available here: www.embedded-computing.com/processing/gti-releases-ai-accelerator-asic-with-integrated-mram.
The GAINBOARD 2803S Inference Accelerator Server will be commercially available in Q1 2019. It will be demonstrated at CES 2019.
For more information, visit https://gyrfalcontech.ai/solutions/.