Ever wondered why we need GPUs or AI accelerators for optimal performance of AI Applications? Ever wondered why a field that existed since the the dawn of twentieth century is only now burgeoning with breakthroughs?
As Kaplan and Haenlein puts it, "AI is a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation”.
Artificial intelligence applications that mimics cognitive functions involves deep neural networks, machine vision, sensor technologies, and machine learning techniques that depends on multi core designs, low precision arithmetics and in memory computing capabilities to work well in both cloud and the edge.
In fields like robotics, autonomous vehicles, drones, decease diagnosis, speech recognition, object or face recognition, capability to vastly expand AI calculations by embracing massively parallel processing power, means near real-time learning, energy efficiency and optimal performance. This is why AI applications requires AI accelerators in its infrastructure design to support various types of computing based on the use cases for sustaining performance during training and inferences.
AI Accelerators
Researchers in scientific fields were probably the first to discover, experiment and adopt GPUs for accelerating AI applications, almost 2 decades ago. Since then, a 'Chip War' has slowly formed along with other technological developments and is now reaching its peak, among cloud companies, Internet giants, chip makers and startups searching for supersonic growth, especially in the fields of artificial intelligence, machine learning and deep learning.
Apart from GPUs (graphics processing unit), other types of microprocessors available in the market currently are FPGA (field programmable gate array), ASIC (application specific integrated circuits), in memory architecture and heterogenous computing. As of today, there is no dominant design that underpins all AI architectures, but NVDIA do seems to command some leadership here while others such as Intel, AMD, WAVE Computing, UK's Graphcore and Israel's Habana Labs are all beginning to ship viable products.
Chinese players...
China has been investing aggressively to produce indigenous AI chips to wean from dependence on US based producers, with ongoing researches underway by major tech corporations, such as Baidu, Alibaba, Tencent and Huawei, apart from government initiatives.
Huawei released several products last year, claiming to be faster than NVDIA with ready adopters in datacenters globally. Another startup called Cambricon released products and is aiming for 30% marketshare in China, setting off the Chinese AI chip industry. One other company worth mentioning here and quite notably working on a customised AI chip is China's leading facial recognition system provider, Sensetime, which is planning to build another three or four supercomputers to process data feeds for facial recognition from millions of cameras nation-wide.
The AI chip leadership momentum will only continue to intensify in 2019....
Intel, Nvidia, AMD and Qualcomm are not the only one competing to produce an omni product that can infiltrate into gadgets, computers, machines and robots. There are others such as IBM, Amazon, Apple, Facebook, Google, Microsoft, Tesla and many more in the Deep Learning space, all making attempts at the design and potentially a universally accepted standard and product.
Facebook recently sent a strong signal by hiring Google's Head of Chip Development, that the social networking firm is serious about building its own semiconductors, joining the likes of Apple, Google, and Amazon.
Cerebras Systems, a startup still in stealth mode hired a top Intel executive, Dhiraj Mallick as its Vice President of Engineering and Business Development. Mallick served as the VP of architecture and CTO of Intel’s data center group prior to this.
What to expect...?
For now, the market for deep learning chips is overwhelmingly dominated by Nvdia graphics processors (or GPUs), which have also been widely used in games and other graphically-intensive applications.
Startups looking to attack this space, has the opportunity to beat the big chipmakers and create a new generation of hardware that will be omnipresent among any AI devices. Think autonomous vehicle, robotics, drones or even a server within a healthcare organization training models for medical problems.
Startups looking to attack this space, has the opportunity to beat the big chipmakers and create a new generation of hardware that will be omnipresent among any AI devices. Think autonomous vehicle, robotics, drones or even a server within a healthcare organization training models for medical problems.
As new products from companies such as WAWE, Huawei, Cambricon, Graphcore and Habana enters datacenters and selected enterprises this year, we might see a flow of special purpose devices being released into the market for AI and deep learning.
AI chip innovation will also aid researchers to make further breakthroughs in various fileds such as Computer Vision, Conversational AI, Natural Language Processing, Reinforcement Learning, Transfer Learning, and General AI. Eventually some businesses and governments from buyer nations, may start to take advantage of the available AI chip offerings and form their own discreet AI clouds for a variety of high profile projects across the organisation for deeper business differentiation and operational excellence as models train faster and learn in realtime.
