Almost every business on the planet is being impacted by the global chip shortage. Consumers are already seeing it directly as planned gadgets and appliances are delayed or in short supply — but they may become much more frustrated when tech's promises of an AI-powered future fall short due to a lack of hardware to operate it on. However, hardware isn't the only method to make these powerful machine learning technologies work.
Over the
last few years, demand for chips of all sizes and powers has steadily
increased, and supply has mainly been able to keep up until the current
production problem. To give you a sense of the scope of this somewhat complicated
tech sector problem, cutting-edge smartphones and servers aren't the only goods
affected by the scarcity. A wide range of consumer products wearables - smart
watches, home automation technology, and automobiles, to mention a few — have
lately been updated to "smart" status and will be impacted as a
result.
The
Current Chip Shortage Solutions
The tech
industry has already begun to take significant steps to address the shortfall.
The obvious thing to take is to invest in existing and new chip production
facilities, which most businesses are already doing. Covid-19, on the other
hand, has disrupted the supply networks that would keep these plants
functioning at normal rates, let alone the expanded ones required to meet
demand.
In reaction
to market instability and political constraints, China's reliable suppliers
have stockpiled and limited their exports, and efforts to make the United
States and others more self-sufficient in electronics production are nowhere
near fulfilment. To put it another way, while investment is necessary to keep
the global chip market afloat, it is insufficient to narrow the gap in the short
term.
A more promising strategy is to accommodate older chip technology, both in terms of production and engineering. When new model inventory runs out, you might consider turning to used automobiles. "Used" semiconductor equipment here refers to manufacturing capacity from past chip generations that is no longer cutting-edge but is certainly better than nothing.
Because of
the large demand for used equipment during the pandemic, device manufacturers
are working on new devices that utilize older chips. This has already helped to
mitigate the effects of the scarcity, but it's a desperate effort for an
industry that, like a shark, must always go forward or perish.
Meanwhile, billions of people use tens of billions of gadgets every day, all of whom may benefit from a more immediate answer: a software solution to a hardware scarcity.
Software Solutions: Smart Compression and Compilation
Unlike hardware, software can be deployed globally at the rate required to maintain the industry's promises of AI-powered cameras, speech and face recognition, augmented reality, and other technologies on track. Until date, the industry has been unable to deploy software as a solution to the chip scarcity and to develop AI models on edge computing devices due to inefficiency.
When it
comes to machine learning, efficient compression and compilation are about much
more than reducing download sizes. It's critical to analyze what aspects of a
working model are crucial to its outcomes in order to lessen the size and power
needs of that model. As a result, smart compression entails "pruning"
the model by deleting layers, filters, or channels without compromising its
accuracy. It also entails "quantization," or reducing precision to
save calculation cycles.
Compilation converts the compressed model's high-level operations to the low-level operations supported by a chip's architecture.
The problem
is that there is no one-size-fits-all solution for completing these critical
tasks. A machine learning model’s complexity must be compressed and assembled
with the target environment in mind. After all, a common chip can be found in a
smartphone, a home automation device, and a scientific equipment, all of which
run distinct operating systems.
The efficiency gains from adapting the compression-compilation design to the exact architecture on which a model is meant to operate can be considerable. Furthermore, popular devices already in the hands of customers can give the real-time AI experience that developers have been pursuing and touting for years. Models ranging from natural language to selfie filters can operate natively faster than they could on specialized hardware, requiring only a regular app install from the user. As a result, the next generation of AI can be implemented without the need for multibillion-dollar infrastructure investments.
For at least
the next three years, there will be a chip shortage. But that doesn't rule out
the possibility of an AI-powered future. Software-based solutions have helped
us get to this stage in machine learning applications, and they may help us go
much further if we use them correctly.
ABOUT
THE WRITER
A
keen observer, love to read geopolitics and investment strategies, writes on the impact of the global semiconductor chip shortage.
No comments:
Post a Comment