Hardware and Software

For AI to succeed, high speed computing is a basic necessity. The data needs to be constantly gathered by a system. Using mathematical and statistical algorithms, decisions have to be made in realtime. Data from any new or unknown situation needs to be synthesized using specific algorithms and made as new model or update existing model, which becomes a part of the constant learning process.

Analysis of Natural Intelligence

An animal running in a certain terrain in a forest will react to jump over an obstacle or turn away using its natural intelligence spontaneously. The AI provides the tools and technology for a machine to achieve or mimic such intelligence. For a machine (robot) to perform such maneuvers operating autonomously, AI algorithms have to -

1. Compute the speed by electronic devices/sensors

2. Constantly make adjustments to speed based on distance computed using radar

3. Avoid obstacles using vision (camera, hardware, IR-sensors, and software)

4. Constantly adjust mechanical devices that perform functions of legs or motion systems based terrain

5. Autonomous tasks such as stop, slow down, maintain speed, accelerate, or jump over obstacles

6. Think and make spontaneous decisions using built-in AI in realtime

At a high level, this lists the complexity involved (intelligence, cognitive analytics) that needs to be incorporated into a robot or autonomous device. The system also has to learn from new situations it encounters on a constant basis.

Graduate School Research

During graduate research on Robotic Dexterous Hand several years ago, I encountered this issue when simulating grasping objects such as a ball by the dexterous hand using software. The software simulation of mechanical grasping of was a success. The dexterous hand was taught (in machine learning terms it is similar to supervised learning) to grab hard balls such as a baseball. The software simulation was mechanically tested using a industrial robotic gripper with three fingers.

The big challenge encountered was when suddenly the object to grasp was switched from baseball to an egg or a ping-pong ball in a demo robot with same robotic gripper with fingers. Without AI (intelligent decision making algorithms) and also due to lack of visual feedback to the robotic gripper, the egg would get squashed. For the dexterous hand or robotic gripper to seamlessly hold objects autonomously without causing damage, requires AI, machine learning algorithms, software and hardware embedded with machine vision.

Due to the computing power of the processors in the 1980s, the actions weren't spontaneous, resulting work-around solutions to embed intelligence or prevent mixing of objects. Thus many industrial robots would perform pre-programmed operations (weak form of AI) with optimum speed.

Computing Power Overview

The table below shows how number of transistors in a single microprocessor have increased steadily over the years resulting in significant improvement in AI. From 1970's to now there has been enormous progress in the computational power due to advancements in memory chip architecture (CPU, VLSI, ULSI, etc.) and technology. The number of transistors started from 2500 in a chip in the 1970s to 67 billion transistors in a single chip in 2023.

The IBM's deep blue supercomputer was able to achieve the level of intelligence to defeat world chess champion Gary Kasparov in a couple of games in a tournament. Autonomous computers and machines continue to evolve with enhancements in AI. As the computing power and other technologies increases, the evolution of AI will be exponential in the future.

One of the CEO of an AI research company was expressing how venture capital companies were ignoring funding of AI and ML in 2010 and around. Going back further to early 1990s, AI was considered as a fantasy and recruiters would advice prospective clients to remove AI from the resumes, whom they were promoting for job placement in client companies. The above graph can be a good reason that the computing power has made AI and ML a thing of the present day.

Use of FPGA In AI

Some search engines used FPGA based hardware in enablement of AI. The FPGA offered high performance/speed, throughput, programmability, flexibility and cost benefits. The development of FPGA hardware was fast. Around 2015, FPGA based hardware was used in search engines such as Microsoft Bing with AI features to enhance search ranking.

Advent of GPU

Initially gaming computer hardware was heavily graphics and video oriented to satisfy rapid graphics/video generation in games. The GPUs were an amazing fit for such applications and supporting hardware. GPU offered Parallel processing Performance that can scale upto supercomputer levels Perform complex arithmetic for AI/Machine Learning - deep learning

The GPU offered many more benefits compared to that of standard CPU, which resulted being a main choice for AI enablement in computing, robotics, image recognition and many more supercomputing systems and applications. The MPP in GPU have made them the best choice for vector/tensor computation that involve massive matrix processing and providing output in nanoseconds or better. Parallel processing of large matrices is a key requirement in neural networks.

Typical NVDIA GPU Chip

The NVDIA GPU chip performance is increasing at exponential rates, from version to version. The performance is complemented by the number of transistors that are packed into a single chip. The NVDIA has several chip technology videos on youtube that show how the high performance, parallel processing, and computational power scaling is accomplised, from one version to the next. This has lead to exponential rate success of NVDIA and AI technology has be complemented by these supercomputing GPU chipsets.

The graphs below show the performance improvement of GPU over time (from 2003 to 2022). The performance improvement is about approximately 7000 and price per performance is about 5600 time higher. This data is from Stanford’s Human-Centered AI group's report on AI [15].

Another graph that shows the exponential improvement of performance scaling of NVDIA GPU chips over time, in FLOPS. The rate at which the number of transistors are being packaged in single chip is pushing the limits of the well known Moore's Law (Moore's law: the number of transistors in an IC will double every two years). Multiple GPU chips are also being merged to attain extreme rate of parallelism and enhance computing power. There are several videos of NVDIA's CEO showing simulation of how it all happens.

Prev   Next

Table of Content

Top