RSS
Логотип
Баннер в шапке 1
Баннер в шапке 2

Facebook: Caffe2

Product
Developers: Facebook
Date of the premiere of the system: 2017/04/19
Branches: Information technologies
Technology: Development tools of applications

Content

2017

Support of Intel Math Kernel Library

On April 24, 2017 the press service of Intel announced expansion of participation in support of own processors in frameworks and their optimization for work with different applied options of training and creation of logical outputs. At the heart of these efforts use of library of the mathematical Intel Math Kernel Library functions (Intel of MKL) which carries out instructions of Intel Advanced Vector Extension CPU.

For April 24, 2017 Facebook and Intel cooperate over integration of functions of Intel of MKL into Caffe2 for ensuring maximum capacity of processors when forming logical outputs.

Properties of Math Kernel Library, (2017)

Performance measures on AlexNet topology when using MKL Intel library in comparison with Eigen BLAS library are given in the table below. In the table the OMP_NUM_THREADS parameter designates number of the physical cores used when processing. At small workloads on creation of logical outputs it is recommended to start each workload in a separate core of the processor and to execute a set of workloads in parallel, on one workload on a core.

  OMP_NUM_THREADS=44 OMP_NUM_THREADS=1
Размер пакета Intel MKL
(изображений в секунду)
Eigen BLAS
(изображений в секунду)
Intel MKL
(изображений в секунду)
Eigen BLAS
(изображений в секунду)
1 173,4 5,2 28,6 5,1
32 1500,2 29,3 64,6 15,4
64 1596,3 35,3 66,0 15,5
256 1735,2 44,9 67,3 16,2

Results of performance of a framework of Caffe2 on AlexNet topology when using MKL and Eigen BLAS Intel libraries. Testing is executed on computers with processors Intel Xeon E5-2699 v4 (with the code name Broadwell) @ 2.20 GHz with two sockets, 22 physical cores on a socket (44 physical cores in both sockets were in total used), 122 GB of RAM DDR4, 2133 of MHz which is disconnected by the option HT under Linux CentOS management 3.10.0-514.2.2.el7.x86_64 3/7/1611, Intel of MKL of version 20170209, Eigen BLAS of version 3.3.2, when using Caffe2 of April 17, 2017.

According to the statement of Intel, results demonstrate high optimization of Caffe2 for work with processors of the company and performance.

Acceleration of Caffe2

On April 20, 2017 NVIDIA and Facebook announced results of collaboration in acceleration of a framework of Caffe2.

Caffe2 allows developers and researchers to create scenarios of the large-scale distributed training and the application of machine learning for end devices.

Providing services on the basis of artificial intelligence on mobile devices should be executed for fractions of a second. For request processing the graphic processors providing data processing rate, the optimized software for problems of deep learning which is capable to implement the potential of a hardware platform are required.

Joint efforts of NVIDIA and Facebook optimized work of Caffe2 using means of deep learning on the platform of the graphic processors NVIDIA. Caffe2 uses NVIDIA — cuDNN, cuBLAS and NCCL SDK libraries – providing training and an output with acceleration on the multi-GPU configurations.

Chart of comparative results of acceleration, (2017)

For April 20, 2017 the partner companies spoke of Caffe2 as a fast, scalable, portable framework of deep learning. It provides almost linear scaling of training of neural networks with acceleration of capacity by 57 times on eight Facebook Big Basin servers, united in network, with 64 NVIDIA Tesla P100 accelerators.

Caffe2

Caffe2 is a set of cross-platform tools for machine learning.

On April 18, 2017 Facebook published the source code of a framework for machine learning of Caffe2. It will allow developers of mobile applications to use AI technologies on mobile devices running iOS and Android, on Raspberry Pi minicomputers.

Caffe2 is tool kit for mobile development. The environment of development will give to users the chance to use sensing technologies of images, natural languag processing and computer vision on the smartphone. Such tasks as a rule are transferred to a remote cloud server, and the results of calculations are returned to the device.

For processing of problems of Caffe2 uses capacities of mobile devices which allow to train simple neural networks. The framework can be used for creation of chat-bots.

The companies worked on acceleration of work of the latest version of a framework NVIDIA Qualcomm Intel, Amazon and Microsoft.