Intel emphasises AI focus

Source: Intel. Baidu’s Jing Wang, Senior Vice President, joined Diane Bryant, EVP and GM, Data Center Group for Intel at the August 2016 Intel Developer Forum to talk about how the Intel Xeon Phi processor makes it easier for Baidu to train their models efficiently.

Intel has turned its focus onto supporting artificial intelligence (AI). The company believes it is the next big wave of compute that will transform the way businesses operate and how people engage with the world. While the company says under 10% of servers worldwide were deployed in support of machine learning last year*, machine learning is the fastest growing field of AI and a key computational method for expanding the field of AI.

Intel's vision of AI is built on the promise of 5G, and makes three assumptions which shape its strategy of creating faster, more intelligent networks:

First, computing will be everywhere and in everything. More than 50 billion things and devices are expected to be connected by 2020, in addition to more than 200 billion connected sensors – all generating massive volumes of data.

Next, compute, analytics and storage capabilities will be distributed into the fabric of the network. Instead of all decisions being made at a central hub, decisions will also be made much closer to the edge of the network. This will be how two autonomous vehicles can sense a collision and communicate directly with each other to prevent an accident.

Third, there will be pervasive connectivity between the things, through the network and to the cloud.

"AI is nascent today, but we believe the clear value and opportunity AI brings to the world make it instrumental for tomorrow’s data centres. Intel’s leadership will be critical as a catalyst for innovation to broaden the reach of AI," said Jason Waxman, Corporate VP, Data Center Group GM, Intel in a blog post. 

At the recent Intel Developer Forum (IDF) in the US, Intel made a number of announcements around AI:

• Commitment to open source with optimised machine learning frameworks (Caffe, Theano). Customers using the optimised version of Caffe are now able to realise up to 30 times increase in performance compared to the mainstream version running on Intel architecture**. Intel will optimise other major machine learning frameworks for Intel architecture by the end of 2016. 
• Optimised the Intel Math Kernel Library (MKL) for common machine learning primitives, allowing deeper access to optimised code through a standard set of APIs at no cost. Intel plans to release the Intel Math Kernel Library – Deep Learning Neural Network, offering an open source implementation of MKL’s deep learning neural network layers. This will facilitate integration and adoption of popular open source deep learning frameworks. 
• Made available the Intel Data Analytics Acceleration Libraries in open source to support deep learning. and libraries. 
• Plans to release a deep learning software development kit (SDK) by the end of 2016.
• Disclosed the next-generation Intel Xeon Phi processor, codenamed Knights Mill, with enhanced variable precision and flexible, high-capacity memory. Knights Mill is focused on high-performance machine learning and artificial intelligence, and expected to be available in 2017.  The Intel Xeon Phi processor family enables data scientists to train complex machine algorithms faster and to run a wider variety of workloads than GPUs. In a 32-node infrastructure, the Intel Xeon Phi processor family can offer up to 1.38 times better scaling than GPUs***.
• The first Intel Silicon Photonics 100G optical transceivers, now commercially available, which will allow enterprises and cloud service providers to use the power of light to move large amounts of information at 100 gigabit-per-second over distances of up to several km over fiber-optic links.
• Announcement of the planned acquisition of Nervana Systems, a leading deep learning provider, bringing together the Intel engineers who create the Intel Xeon and Intel Xeon Phi processors with Nervana’s machine learning experts to accelerate progress in AI

According to Intel, its Xeon processor E5 family is the most widely deployed processor for machine learning inference, with the added flexibility to run data centre workloads. Combining these processors with Altera's Arria 10 field programmable gate arrays (FPGAs) delivers excellent performance per watt and the ability to reconfigure the device to manage various workloads.

The Intel Scalable System Framework further offers reference architectures and designs that enhance technology interoperability and reduce deployment complexity, offering a path to broad adoption of distributed deep learning algorithms and reduction in time-to-model.

Intel also stressed that graphics processing units (GPUs) - notably from rival NVIDIA - are not mainstream. "While there’s been much talk about the value of GPUs for machine learning, the fact is that fewer than 3% of all servers deployed for machine learning last year* used a GPU," noted Waxman.

"It’s completely understandable why this data, coupled with Intel’s history of successfully bringing new, advanced technologies to market and our recent sizable investments, would concern our competitors. However, arguing over publicly available performance benchmarks is a waste of time. It’s Intel’s practice to base performance claims on the latest publicly available information at the time the claim is published, and we stand by our data."

Interested?

Read the TechTrade Asia blog post about NVIDIA's deep learning capabilities

View the infographic on artificial intelligence (PDF)

*Source: Intel estimate

**Up to 30x software optimisation improvement claim by Intel is based on customer CNN's training workload running 2S Intel Xeon processor E5-2680 v3 running Berkeley Vision and Learning Center Caffe + OpenBlas library and then run tuned on the Intel Optimized Caffe (internal development version) + Intel Math Kernel Library.

***Up to 38% better scaling efficiency at 32-nodes; claim based on GoogleNet deep learning image classification training topology using a large image database comparing one node Intel Xeon Phi processor 7250 (16 GB, 1.4 GHz, 68 cores) in Intel Server System LADMP2312KXXX41, DDR4 96GB DDR4-2400 MHz, quad cluster mode, MCDRAM flat memory mode, Red Hat Enterprise Linux 6.7, Intel Optimized DNN Framework with 87% efficiency compared to unknown hosts running 32 NVIDIA Tesla K20 GPUs each with 62% efficiency (Source: http://arxiv.org/pdf/1511.00175v2.pdf showing FireCaffe with 32 each NVIDIA Tesla K20s - Titan supercomputer - running GoogLeNet at 20x speedup over Caffe with one K20 each).