AI Hardware
Infrastructure
Report UAE

The United Arab Emirates (UAE) government is in a phase of rapid growth in terms of the adoption and development of AI services.

Machine Learning,
Computer Vision and NLP

Executive Summary

The branches of AI including machine learning, computer vision and NLP have seen unprecedented growth in the past 10 years. The rapid development of these technologies was a result of the growth in computation power, memory, and their affordability. In addition, there is a unanimous agreement from governments, industries and academia across the world on the fact that developing AI holds significant benefits to society and the economy. The market size for AI is rapidly expanding and is expected to grow even faster. In anticipation for that growth, the UAE government took the initiative by appointing the first ever minister of Artificial Intelligence in October of 2017. This was followed by releasing the UAE AI strategy which outlines a vision for the UAE to be one of the leading nations in AI by the year 2031.

With such an ambitious vision comes a number of initiatives designed to accelerate up skilling talent, developing expertise, setting up a governance framework for AI and finally building up an infrastructure capable of supporting the growing demand for computational power for both developing and deploying AI solutions in the UAE. Being a leader in AI requires pushing the envelope with regards to research. However, the algorithms and data are growing in complexity which requires more and more computational power than ever. For example, an autonomous vehicle can generate on average over a petabyte of data per month if operated for only 4 hours a day. Storing, processing and learning from this data is not possible on consumer grade hardware. In fact, working on such applications require specialized hardware generally referred to as High-Performance Computing platforms (HPCs) or also known as Super Computers, which are specialized machines that are an aggregate of many high-performance processing, memory units and storage units.

Given that computational power is a key enabler of AI, the UAE has developed this report as an initial step to fully understand and compile a comprehensive understanding of the existing hardware infrastructure used for AI applications in the UAE. Thus, this report is set to define a baseline for the current state of the hardware infrastructure for AI in the UAE, then compare it to other leading nations in the field. The report will also highlight the main characteristics of a typical AI high performance computing cluster.

40 Government, Private and Academic entities

Brief and Methodology

This report is the initial step to understand the current state of the AI hardware infrastructure in the UAE. The data presented in this report is a result of a survey that was carried out with over 40 government, private and academic entities as well as meetings with key players to discuss the details of their infrastructure. The survey covered the following:

Hardware specifications of the platform used for AI applications
- CPU count and specifications
- GPU cont and specifications
- Total system memory
The benchmark used to measure the performance
Whether it is physically located in the UAE
How the entity utilizes these supercomputers
- The problems these systems solve
- Whether it is available for the general community to use or not

_{CPU: Central Processing Unit}
_{GPU: Graphics Processing Unit}

Due to the difficulty of standardising the performance metrics of different hardware specifications, benchmarks and the data reported by these entities, the report normalised and approximated certain performance metrics based on publicly available information and reports (We plan to standardise the main performance metrics for the next iteration of the report to be released in April 2021).

Data Infrastructure,
Network and Hardware Infrastructure

Introduction

There are a number of key enablers to build a successful and efficient ecosystem to support innovation and research in AI. Some of the key enablers come in the form of the infrastructure to support entities developing, deploying and researching AI. The main pillars of the AI infrastructure can be categorized into three main items

Data infrastructure: Generally, refers to the availability of data and the high-performance storage platforms that synergizes with AI applications.

Network: Specialized high-performance networking systems that connect the compute servers amongst themselves and to storage units.

Hardware infrastructure: The computing platforms and computer chips that accelerate the process of training and deploying AI applications as well as supporting a large amount of memory.

benchmark: A software or set of tasks that is run by computers to evaluate their performance.

The trend in AI development is leaning towards more complex and larger models that require large computational capabilities. In fact, it is becoming more difficult to train and deploy AI models on home desktops. Thus, this report focuses on the compute infrastructure as a key enabler for pushing the boundaries of AI.

The process of training AI algorithms is a very complex task that requires a significant amount of domain expertise. However, in its simplest form, training a machine learning model is simply the process of feeding a large amount of data to a set of algorithms. A set of simple computations are repeated numerous times where in every cycle of this process, the algorithm slowly converges towards more accurate predictions (learns to better perform the task). Thus, in its nature training algorithms is a very repetitive task and is highly parallelizable. This makes it synergetic with some types of computer hardware.

The hardware components designed for accelerating the learning process of AI algorithms are very different from typical data center hardware designed for service centers and cloud platforms. Compute units typically emphasizes a large number of processing units (by coupling CPUs and GPUs) to achieve very high processing capabilities measured in number of operations per second (FLOPS). In loose terms, the higher the FLOPS count, the more capable the system is for performing the operations needed to train and develop AI algorithms. When designing or upgrading hardware infrastructure, it is important to make a choice between expanding the CPU count or GPU count of the system. These two components are similar in that they perform the processing tasks. However, the similarities end there as these two components are designed with different purposes in mind.

Central
Processing
Unit
vs
Graphical
Processing
Unit

CPU vs GPU for AI Workloads

CPUs (Central Processing Unit) are the main processing units of any computing machine, these components are designed to be a general-purpose processing unit that are able to execute a wide variety of workloads. Building a platform that is mainly composed of CPUs adds a large degree of versatility to the platform making it suitable for data center and generic cloud workloads. This is mainly due to its design choices making it well suited for serial tasks by having limited number of more capable cores. Furthermore, CPUs can also perform AI algorithm training tasks albeit at a much lower performance than preferred. CPUs can also be provisioned/assigned in finer grained resources to perform smaller tasks. Overall, expanding a platform CPU count is a safe choice to steadily increase its processing performance (given that there are no other bottlenecks in the system).

On the other hand, GPUs (Graphical Processing Unit) play a different role in a system, these components are not as versatile as CPUs. However, they make up for it with a much higher performance in very specialized tasks such as 3D rendering and training deep learning models. GPUs are composed of a very large number of computational units (cores) specialized in performing simple mathematical tasks rapidly in parallel which is very suited for training machine learning models. In fact, a single GPU can have thousands of cores compared to tens of cores on a CPU, which makes GPUs vastly superior to CPUs in AI applications. However, the cost of a GPU is significantly higher than a CPU and it remains a highly specialized component that is not suitable for all workloads.

The UAE is a growing leader

Existing AI Hardware Infrastructure in the UAE

The UAE is a growing leader when it comes to AI infrastructure. In fact, the UAE (at the time of writing) has the 36th most powerful high-performance computer in the world5. The UAE’s AI infrastructure ecosystem is dominated by the private sector owning 89% of the total AI related processing power in the UAE. The academic community comes in second and is home to some large-scale state of the art systems, for example NYUAD hosts “Dalma”, a supercomputer composed of more than 870 CPUs and 30 GPUs. These are not the only AI systems in the UAE according to our survey. On the contrary, our survey identified more than 10 high-end systems owned by the government, academia and the private sector. Below is a compiled summary of their combined resources.

Distribution of Computation Resources in the UAE

The UAE
currently ranks
36 globally

International Position

The international benchmark and ranking for supercomputers focus mainly on the total processing power of the system. The most notable list containing information on the most powerful supercomputers in the world is “Top500”³ list. The UAE currently ranks 36 globaly with the “Artemis” high performance computing cluster owned by G42.

Tianhe-2 The sixth most powerful systems in the world - China-

Country Ranking per number of systems on the TOP500

Countries having more supercomputers on the top500 list will rank higher

Ranking of countries on the top500 list ordered by country’s fastest supercomputers

Summit Supercomputer The second most powerful system in the world -USA-

Repeating countries were filtered out.

Key Participants

Group 42 is an Abu Dhabi based artificial intelligence (AI) and cloud computing company that owns the 36th “Artemis” and the 304th “POD3” most powerful supercomputers in the world.

G42>>

RTA: The Road and Transportation Authority of Dubai hosts a CPU and GPU hybrid system which is utilised for conducting Transport and Corporate related analytics and projections.

RTA >>

ADDA: Abu Dhabi Digital Authority is currently in the process of building one of the largest local government platforms to support the development of their projects and services.

ADDA GOV AE >>

Morohub: A Dubai based private entity that offers data center service. Its “Moro Cloud” is composed of both CPU and high-end GPU units.

MORO HUB >>

ADNOC currently houses a powerful supercomputer that is performing scientific oil reservoir simulations and seismic processing.

ADNOC >>

UAEU: UAEU is based in Al Ain with over 13,000 students. The university owns 3 high performance platforms that are used by various departments for research.

UAEU >>