Chip Based Artificial Intelligence Computing

When AI computing was still in its infancy in the 1990’s through 2000’s, enthusiasts, students, researchers and developers relied heavily on CPU’s and GPU’s for processing. The latter was popularized by Nvidia in 1999, who marketed the GeForce 256 as "the world's first GPU" (Graphics processing unit - Wikipedia). It was thought to be the ultimate solution for processing big data and for other computational requirements. The CPU and GPU architecture must also be supported with their corresponding software: drivers and API’s for interfacing and optimization.

Overcoming Latency and Throughput Challenges

Even today’s University Computing Laboratories generally have been dependent on the said architectures let alone AI applications and algorithms suffered due to latency of compiling executive files and processing throughput. When these technical problems are naturally occurring during laboratory experiments, the choice of algorithm, optimization of programming codes among other reasons could have been the victims for revision or eventual change. So much wasted time for exploratory works in order to address the issue of latency, throughput, and security.

When AI is applied in manufacturing industry like real-time intelligent cement processing that requires say, robotic material scanner and classifier, sensory motors for material flows, transmitters, temperature sensors, real-time digital counters for huge silos, among others, such issues on latency and throughput will not do anything strategic for the entire real-time process. It would result to low quality of products that cannot pass quality control and evaluation to be performed by the “would be clients” affecting product marketability. With this, computer engineers and scientists would rather resort to wired architecture than software-based. There are no programming lines of codes written to execute commands but “hard-wire programming” or logic circuits comprising millions of transistor gates are designed to perform specific tasks in the production lines. For instance, when the temperature sensor detects 90-degree C along the line, a sensor for cool air mechanism activates opening the valve blowing cool air into the affected area needing it and activates another sensor to sprinkle water for fast dropping down of temperate. If the temperature refuses to drop normally, the logic circuit shuts down that portion of the line and activate another sensor for sustained alarm catching the attention of engineers and technicians on duty.

The Drawback and Opportunity of Special Purpose Machines in AI Applications

In above scenario, if the wired commands had been implemented thru software, chances are: it executes properly the commands or the system is compromised or hacked purposely to set the plant on fire, thereby burning the facility. Chip-based AI computing addresses security concerns, throughput and latency issues for all applications. Much more, with the advent of nano technology, designing and the construction of a very large-scale integration (VLSI) chips is no longer a problem. Unarguably, if AI is implemented through nano technology- chips, response is amazingly in nano second

There is of course, a drawback and opportunity in this argument. Multiple machines of specific application could be thriving. Many in the computing industry would opine such concept is contrary to the tenet of flexibility and slightness/diminutiveness of machines. Scenarios like you buy a special purpose machine for Photoshop, Corel draw, python language, juice production line are old approaches like that of ENIAC’s (Electronic Numerical Integrator And Computer) as the world's first general-purpose computer (https://www.techtarget.com/) if not of “Charles Babbage’s Analytical Engine (he worked on it until his death in 1871) (https://science.howstuffworks.com/), (Analytical Engine | Description & Facts | Britannica)”

Not until in 1958 when Jack Kilby of Texas Instruments patented the principle of integration, created the first prototype ICs and commercialized them (Invention of the integrated circuit - Wikipedia), the idea of chip-based AI was not conceived yet. In the early 1970s, MOS integrated circuit technology allowed the integration of more than 10,000 transistors in a single chip. This paved the way for VLSI in the 1970s and 1980s, with tens of thousands of MOS transistors on a single chip (later hundreds of thousands,   then millions, and now billions) (Very-large-scale integration - Wikipedia). Today, with the real-time requirements of reduced latency, high throughput and high security, the VLSI or even Ultra Large-Scale Integration (ULSI) and the utilization of nanotechnology have rendered designing and constructing chips of even up to 2048 bits for specific AI application highly feasible in all respects including diminutiveness of actual machines.

For clarity, this concept is strategic for any critical AI applications demanding low latency, high throughput, and high security. Anyway, one can order from a chip manufacturer based on requirements. All intelligent algorithms such as fussy logic, artificial neural net (ANN), and other deep learning approaches can very well be integrated into during the very large-scale integration designing phase of the chip-based process       

The foregoing argument: the world of AI operates the system with confidence while hackers will experience a ruinous-heap starvation before they may be able to design a new algorithm to interfere with the system..