引用 ・Unlike usual FPGA-based solutions, Zebra works on the user-defined neural networks as it would on other processing engines like CPU and GPU. ・Zebra users don’t have to learn new languages, new infrastructures or new tools, it works with Caffe, MXNET, TensorFlow, etc. ・Zebra supports 16-bit fixed point, 8-bit fixed point, and more. ・Zebra includes FPGA contents and SW stack. It is simply like replacing a GPU by a FPGA. There is no long wait for FPGA compilation, Zebra runs immediately. ・Zebra runs on FPGA boards typically using less than 40W. ・Zebra can run on any FPGA. With latest FPGA generation offering higher calculation bandwidth, Zebra users run faster the same content byjust installing a new board.
引用 Zebra | Version 16-12 Supports TUL-KU115 (based on Xilinx KU115). Performance of AlexNet inference withint16: over 1000 img/s. Power: below 40W Perf/W : approx. 30 img/s/W. Caffe infrastructure. Demonstration on demand Zebra | Version 17-06 large datacenter deployment, or in-house PCIe boards. Supports all NN similar to AlexNet and GoogLeNet. Caffe & MXNET infrastructures. Expected AlexNet inference performance: int16 over 2,400 img/s, int8 over 4,500 img/s. Supports KU115-based & VU9P-based boards (VU13P-based boards upon availability). Power targeted: below 40W, 100 img/s/W. Try it on AWS Marketplace (ES2 F1): just look for the Zebra.
FPGAなので定期的なバージョンアップ(これによると6ヶ月)があるのね。
4週間でAWS EC F1に移植したみたい。その後4週間でチューニングを行ったと。
ユーザ側ではFPGAでの開発の必要は無い模様。。。