2017 PRC Spring AI Technical Workshop
会议时间:2017年3月28日(8:30-16:45)
签到时间:8:30-9:30
会议酒店:北京金隅喜来登酒店3F宴会厅
酒店地址:北京市东城区北三环东路36号
INTER AI HARDWARE PLATFORM
- inter nervana portfolio
inter xeon processors
inter xeon phi processors
inter xeon processors + lake crest
inter xeon processors + fpga roadmap
crest family (nervana)
alter fpga
inter xeon phi processors
nervana processor
inter xeon processor sp 25up to 400X caffe performance on intel
INTER AI SOFTWARE STACKS
- github intel/caffe intel/theano
Nervana Solutions
customer Neon Solutions
Customer Models Neon Models
Customer Algorithms Neon Deep Leaning Functions
Hardware Agnostic Inter Nervana Graph Compiler
Hardware efficient Transformersinter nervana al academy
http://software.inter.com/ai
frameworks,tools and libraries
BIGDL : A distributed DL library form Apache Spark
DAAL : open source library
inter python distribution powered by anaconda
总结
- 硬件: FPGA 相同计算性能时和gpu相比,功耗低,成本低,功耗比约1:3(ppt上给的数椐fpga130w,gpu350w)。结合inter xeon phi 相关处理器使用,此处理器增加了许多能提高深度学习相关计算的新指令集。
- 软件:intel 对caffe https://github.com/intel/caffe 和 tensorflow针对自己的芯片架构做了修改优化。 intel 4亿美金 收购的 初创深度学习公司Nervana 开源的ngraph https://github.com/NervanaSystems/ngraph