NovaFinance.AI 
By: Tymur Chalbash
Artificial Intelligence (AI) has moved beyond science fiction, rapidly permeating every aspect of our lives, from recommendation systems to autonomous vehicles and medical diagnostics. However, this explosive growth wouldn’t be possible without relentless progress in hardware, particularly in the development of increasingly powerful and energy-efficient chips. New generations of processors, graphics accelerators, and specialized AI chips are having a colossal impact on the speed of AI technology development and the breadth of its adoption, opening doors to solving problems that seemed unattainable just yesterday.
The Crucial Role of Computational Power
One of the key factors driving progress in AI is computational power. Training modern deep neural networks requires processing enormous amounts of data and performing trillions of floating-point operations. Early stages of AI development were significantly limited by the capabilities of existing processors, which slowed down both research and the practical application of developed models.
The emergence of Graphics Processing Units (GPUs), initially designed for processing graphics in video games, was a real breakthrough. Their parallel architecture proved to be ideally suited for performing the matrix operations that underpin many machine learning algorithms. As Yann LeCun, one of the pioneers of deep learning, notes: “GPUs gave us the ability to train much deeper and more complex models in much less time. This became one of the key factors in the success of deep learning.” [1]
A prime example is the development of Convolutional Neural Networks (CNNs) for image recognition. Training early CNNs on central processing units (CPUs) took weeks and months, while the use of GPUs reduced this time to days or even hours. For instance, training the AlexNet model, a revolutionary CNN from 2012, took 5-6 days on two NVIDIA GTX 580 GPUs. [2] In comparison, a contemporary CPU might have taken several months. This, in turn, accelerated research in computer vision and led to the advent of technologies like facial recognition, automated medical image processing, and self-driving cars.
The Rise of Specialized AI Accelerators
However, the demands for computational power continue to grow as AI models become more complex and data volumes increase. This is why the microelectronics industry isn’t standing still, developing new generations of chips specifically optimized for artificial intelligence tasks.
One of the most promising areas is specialized AI chips (AI accelerators), such as Google’s TPU (Tensor Processing Unit), Huawei’s NPU (Neural Processing Unit), and others. These chips feature architectures designed from the ground up for the efficient execution of operations characteristic of machine learning, such as tensor computations and sparse operations.
According to Sundar Pichai, CEO of Google: “TPUs were designed to make machine learning accessible to everyone. They have allowed us to significantly accelerate the training and inference of our most advanced models, which, in turn, has spurred the development of products like Google Translate, Google Photos, and AlphaGo.” [3]
A vivid example of TPU efficiency is the training of the BERT language model. While training BERT on powerful NVIDIA V100 GPUs took about 3-4 days, on Cloud TPU v3, this time was reduced to one day, and on Cloud TPU v4 (a newer generation), to a few hours. [4] This dramatic reduction in training time allows researchers to test new ideas faster and iteratively improve models.
The use of specialized AI chips not only significantly increases the speed of model training and inference but also reduces power consumption. This is especially crucial for deploying AI solutions on mobile devices and in cloud environments. For example, the high energy efficiency of these chips allows for the creation of more compact and long-lasting smartphones with machine learning features, such as natural language processing and real-time object recognition. For instance, the Neural Engine in recent iPhone models can perform trillions of operations per second (TOPS), enabling complex AI tasks to be processed directly on the device, without requiring constant cloud connectivity and saving battery life. [5]
Software, Frameworks, and Expanding Applications
Beyond increasing computational power, new generations of chips also impact the speed of AI development by improving tools and frameworks. Chip manufacturers actively invest in developing software, libraries, and APIs that make it easier for developers to utilize new hardware capabilities. This allows researchers and engineers to focus on developing models and algorithms themselves, rather than spending time on low-level code optimization for specific hardware.
For example, NVIDIA developed the CUDA platform, which provides developers with a convenient interface for programming their GPUs. [6] TensorFlow and PyTorch, two of the most popular deep learning frameworks, have built-in support for CUDA and other hardware accelerators, enabling automatic utilization of available computational resources. [7]
Furthermore, advancements in chips contribute to the expansion of AI’s application scope. More powerful and energy-efficient hardware makes it possible to implement complex AI models in new areas such as autonomous transport, personalized medicine, and industrial automation.
As an illustration, consider the development of self-driving cars. Processing data from numerous sensors in real-time and making instantaneous decisions requires immense computational power. New generations of chips specifically designed for the automotive industry can provide the necessary performance and safety for autonomous driving systems. NVIDIA Drive Orin chips, intended for autonomous driving, can achieve performance of up to 254 TOPS, allowing them to process data from dozens of cameras, lidars, and radars simultaneously, forming a complete picture of the surroundings and making decisions in milliseconds. [8] As Elon Musk, CEO of Tesla, notes: “Developing specialized chips is critically important to achieving true autonomous driving. We are constantly working to improve our hardware to make our cars even safer and smarter.” [9]
In medicine, thanks to the growth in computational power, AI is actively applied to analyze medical images. For instance, for early detection of cancerous tumors, where speed and accuracy are critical. Systems using modern GPUs can process thousands of MRI or CT scans in minutes, identifying anomalies that may be unnoticeable to the human eye. [10] This significantly accelerates diagnosis and allows for earlier treatment.
Conclusion
Thus, new generations of chips are a fundamental factor determining the speed of artificial intelligence development and adoption. The increase in computational power (from tens of GFLOPS for CPUs to hundreds of TFLOPS and even PFLOPS for modern GPUs and AI accelerators), energy efficiency (reduced power consumption per computation), and the emergence of specialized architectures open new possibilities for solving complex problems and expand the scope of AI’s application in all aspects of our lives. Continuous progress in microelectronics will continue to play a key role in the further advancement of artificial intelligence, bringing us closer to a future where intelligent systems become an integral part of our daily lives.
Sources:
[1] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444.
[2] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Proceedings of the 25th International Conference on Neural Information Processing Systems (NIPS 2012), Vol. 1, pp. 1097-1105.
[3] Pichai, S. Google Cloud Blog. (2016). Cloud TPU: Custom chips for machine learning. Available at [Google Cloud Blog archive for 2016 or official announcements].
[4] Google Cloud. Google AI Blog. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
[5] Apple. (Official product pages and developer documentation for iPhone/A-series chips).
[6] NVIDIA. (Official CUDA platform documentation). Available at: https://developer.nvidia.com/cuda-zone.
[7] TensorFlow Documentation. (N.d.). GPU support. Available at: https://www.tensorflow.org/install/gpu. PyTorch Documentation. (N.d.). CUDA semantics. Available at: https://pytorch.org/docs/stable/notes/cuda.html.
[8] NVIDIA. (Official product pages and press releases for NVIDIA Drive Orin). Available at: https://www.nvidia.com/en-us/self-driving-cars/drive-platform/drive-orin/ (Note: Link might evolve).
[9] Musk, E. Tesla’s Autonomy Day presentations or quarterly earnings calls where custom chip development is discussed.
[10] Academic Research / Clinical Trials. (Numerous studies on AI in medical imaging). Esteva, A., et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115-118.
