AWS and Intel: Reinvent the Future of Cloud

Cloud took Las Vegas by storm last week with the annual Amazon Web Services (AWS) re:Invent conference. This was the sixth year Intel sponsored this cornerstone cloud computing event and we have showcased the full range of our collaboration with AWS, from edge to cloud, spanning an array of applications such as artificial intelligence (AI), high-performance computing (HPC), and the Internet of Things (IoT). I’d like to recap some highlights of the Intel and AWS collaboration over the last year and highlight some of the innovations we showcased at re:Invent 2017.

Relentless Pursuit of Performance and Efficiency

One of the biggest moments for Intel this year was the launch of Intel® Xeon® Scalable processors on July 11. AWS was a key customer for this revolutionary platform, and we were very pleased when Dr. Matt Wood, Amazon’s General Manager of Artificial Intelligence, supported our launch talking about AWS EC2’s new compute-intensive C5 instance, powered by Intel Xeon Scalable processors. He emphasized the importance of C5 for HPC, machine learning, and AI applications, and stated that the computational power of the Intel Xeon Scalable processor lets AWS customers use more data and machine learning to create innovative new products and experiences. Dr. Wood discussed how Intel and AWS collaborated on the optimization of C5’s deep learning engine with Intel® Math Kernel Library (Intel® MKL), increasing AI inference performance by over.

Through Intel’s long collaboration with leading Cloud Service Providers (CSPs) such as AWS to drive cloud innovations, delivering the highest level of cloud infrastructure efficiency has always been a priority because at cloud scale, every ounce of efficiency matters. Across generations of Intel technology and product advances, we have consistently enhanced performance/$TCO (dollar of Total Cost of Ownership), and along with our CSP partners passed these savings and benefits to end users across a wide range of cloud workloads. Let me give you an example that we highlighted at this re:Invent that showcased the efficiency gain that the new AWS C5 instances have over C4 instances family for High-Performance Computing (HPC) workloads. For a select set of HPC application workloads, c5.18xlarge instances demonstrated performance gains of 1.74X to 4.66X compared to c4.8xlarge1 instances. When we take into consideration of price differentials between C5 and C4, C5 still demonstrated more desirable price-performance ratio for the majority of HPC application workloads1.

Relative total cost of ownership (TCO) of HPC workloads comparing c4.8xlarge instances to c5.18xlarge instances 1
Relative performance of HPC workloads comparing c4.8xlarge instances to c5.18xlarge instances 1

The momentum continued with the M5 announcement on November 28th by Jeff Barr, which is historical considering that M instances, specifically the m1.small instance, were what started AWS. The new M5 general-purpose instances are based on Custom Intel® Xeon® Platinum 8175M processors running at 2.5 GHz and designed for today’s highly demanding workloads. Customers can use M5 instances to run web & app servers, host enterprise applications, support online games, and build cache fleets more efficiently. According to Jeff, M5 will deliver 14% better price/performance than the M4 instances on a per-core basis. Applications that use the AVX-512 instructions will crank out twice as many FLOPS per core. A new size with 96 vCPUs was made available at the high-end, giving customers even more options.

Acceleration of New Cloud Service Creation

Another collaboration between Intel and AWS that helped AWS bring new services and capabilities to the cloud was around BigDL, an open-source, distributed deep learning framework for Apache Spark that bridges big data and deep learning.  As deep learning gains traction in applications such as recommendation engines, voice and speech recognition, and image and video recognition, many of AWS’s analytics customers want to incorporate deep learning with massive amounts of data in Apache Spark, rather than feed it into a separate deep neural networks infrastructure to train a model. Integrating big data and deep learning can significantly increase efficiency in the end-to-end data analytics pipeline. BigDL is implemented as a library on top of Spark, and run directly on top of existing Spark or Hadoop clusters, such as AWS Elastic Map Reduce (EMR) service and integrates with popular AI frameworks like Apache MXNet and TensorFlow. By making BigDL AWS Machine Image (AMI) available on AWS MarketPlace, AWS customers can be on their way to a smoother deep learning-powered analytics journey.

Similar to HPC workloads, deep learning training workloads running on BigDL framework also saw a significant performance and price-performance enhancement between C4 and C5. Across a range of deep learning models including VGG cifar-10, ResNet-50, and ResNet-152, C5 delivered up to 2X speed-up in terms of image throughput/second, and more desirable price/performance3.

Relative throughput performance of images/sec comparing c4.8xlarge instances to c5.18xlarge instances 3
Cost savings for model training by comparing c4.8xlarge instances to c5.18xlarge instances 2

Ubiquitous Compute from Edge to Cloud

At Intel, we work hand-in-hand with our cloud customers such as AWS to build the future now. With a comprehensive, end-to-end portfolio of processors, accelerators, devices, and software, only Intel can address the computing needs from edge to cloud, providing users a seamless computing experience for emerging IoT, autonomous driving AI and analytics applications.

For developers who yearn for cutting-edge technologies and coolest gadgets, you must check out DeepLens that AWS CEO Andy Jassy announced at re:Invent. You can get hands-on with deep learning with this fully programmable video camera, tutorials, code, and pre-trained AI models, all in a box designed to expand deep learning skills. Powered by an Intel® Atom® X5 processor with embedded graphics that support object detection and recognition, DeepLens uses Intel-optimized deep learning software tools and libraries (including the Intel Compute Library for Deep Neural Networks, Intel clDNN) to run computer vision models directly on the device for real-time responsiveness. It can also seamlessly make use of the cloud for more compute-intensive higher-level processing. For example, you can do face detection on the DeepLens and then let Amazon Rekognition take care of the face recognition.

No matter whether you were in Vegas for re:Invent or not, you can always visit the Amazon Services & Intel page to find out the latest joint innovations between AWS and Intel.

1 See https://www.intel.com/xeonconfigs for configuration details "Footnotes 21". Testing conducted on HPC applications and workloads comparing AWS c4.8xlarge vs. c5.18xlarge instances. Testing by Intel. 
2 TCO model: Given Cost: C5=$3.06/Hr, C4=$1.59/Hr. Cost Ratio C5/C4=1.92.TCO Ratio between C5/C4= Cost Ratio(C5/C4) / Perf Ratio (C5/C4).  C5/C4 perf ratio needs to be > 1.92 to be cost efficient
Cost model based on current EC2 pricing:
(https://aws.amazon.com/ec2/pricing/on-demand )
c5 cost=$3.06/hr, c4 cost=$1.59/hr
Workload cost based on pro-rated hourly rental costs for all instances used.Notices & Disclaimers
3 See https://www.intel.com/xeonconfigs for configuration details "Footnotes 23". Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations, and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For complete information visit http://www.intel.com/benchmarks.