Sergey Kostrov wrote:
- Intel 7th Generation CPU ( Xeon Phi / Knights Mill )
- Intel Stratix 10 FPGA 14nm Tri-Gate process ( Sampling Q4 2016 / ~30 billion transistors )
- Intel Silicon Photonics
- Intel Systmem Studio for Microcontrollers
- Intel Instrumentation and Trace API ( IIT API )
- Intel MKL DNN
- Intel Deep Learning SDK
- Intel Performance SnapShot
- Intel SNAP
Types of FPGAs:
- Altera FPGA
- Microsoft Catapult V1 Card FPGA
- Google TPU Card ASIC
- VeriLog / VHDL
Types of processing:
- I/O-bound ( aka Disk-bound )
- DC - Data Center
- HBM - High Bandwidth Memory
- HDF5 - Hierarchical Data Format v5
- NN - Neural Network
- DNN - Deep Neural Network
- ML - Machine Learning
- DL - Deep Learning
- Capacity / Low Latency / Power Consumption
- CentOS v7.2
- Code-Based Algorithm
- Data-Driven Algorithm
- Numerical Library PETSC
- Rack Scale / QCT / Inspur
Thanks for sharing your notes. I feel :
Intel Deep Learning SDK
Intel Performance SnapShot
Intel Aero Ready-to-Fly Drone
Intel Project Alloy VR headset and
Intel Joule Board
are the hot topics for Intel Dev community.
Moreover, wanted to point out the link you shared - http://github.com/hfp/libsxmm , lands on a 404 error page.
Janko (Bayncore) wrote:
Hi there, the correct link to matrix library on GitHub is https://github.com/hfp/libxsmm (xs switcheroo) and it is actually highly optimized and recommended.
That's the correct URL. Thank you Janko!
Related to the IDF'16 topic, I also want to point out that we changed our summary:
Library targeting Intel Architecture (x86) for small dense and sparse matrix multiplications as well as small convolutions.
The "small convolutions" may help with emerged workloads such as deep learning and more specifically Convolutional Neural Networks (CNN). The latter are pushing the limits of today's hardware, and one of the expensive kernels is a small convolution with certain kernel sizes (3, 5, or 7) such that calculations in the frequency space are not the most efficient method when compared with direct convolutions. LIBXSMM's current support for convolutions aims for an easy to use invocation of small (direct) convolutions, which are intended for CNN training and classification.