AI Chips are not CPU's
AI needs a
different kind of specialized chip. So no, it is not just a faster CPU (Central
Processing Unit). Rather, it is a specialized variant of CPU called GPU
(Graphical Processing Unit). GPU’s were originally designed ages back for
one purpose – video games! What?! The “G” was for Graphical, remember. In a
video game, the screen (1) has to be rendered at insane speeds, (2)
how to render each pixel involves a combo of RGB values (among other things)
which have to be crunched via matrix multiplication, and (3)
rendering of pixels can be done in parallel, not in sequence the way
traditional computing is done.
Decade later, the
raw data for AI needed just such a processor – insane speed, matrix
multiplications, and parallel processing. Traditional CPU’s were not designed
for that, but these were exactly what GPU’s already did. Ergo, the company that
started off decades back with GPU’s for video games pivoted to AI chips and is
the world’s most valuable company today – Nvidia. Yes, it’s more
valuable than Apple, Microsoft, Google and Amazon…
~~
At the beginning,
there was Silicon Valley. So named because semiconductor companies started off
there, most famously Intel. Being among the first, Intel didn’t have to just
design the IC’s, it also had to manufacture them (the chip factories are called
foundries). Foundries are extremely expensive; so the ability to design
better chips was worthless if you didn’t have a foundry to manufacture those
chips in.
This situation led
to the creation of TSMC, the Taiwanese chip manufacturing giant of
today. Basically, it was a common chip manufacturing factory for anyone and
everyone. Anyone could now design an IC, and TSMC would manufacture for them.
Over time (decades), the chip industry split into specialists – IC design
companies with no foundries; and TSMC as the one common manufacturer for all of
them (except Intel, which remember, had its own foundries). This set off a
virtuous cycle at TSMC – since everyone came to them (other than Intel), they
had huge volumes, leading to efficiencies of scale. They were not tied to any
one company, though some like Apple had outsized priority. TSMC could afford to
keep investing in smaller and smaller chip size, better equipment etc since
there were so many IC designers out there.
Today, only
TSMC can manufacture the highest end chips on the planet. Including those AI
chips of Nvidia. If TSMC were to go down tomorrow for any reason, including
yes, China invading Taiwan, the world would face an immediate and unsolvable
chip problem across all ends – low, mid and high. TSMC, therefore, is called
the Silicon Shield of Taiwan – the US and the world cannot let China
invade Taiwan because everyone needs those chips from TSMC.
~~
All of this is the
cause of the geopolitics of chips today. The best foundry is in Taiwan (TSMC),
the top AI chip designing firm, while American, is the baby of a Taiwanese
immigrant (Nvidia), and China is the world’s largest market for more and more
things, including those AI chips.
Hardly an
appealing situation for America. Attempts to prevent sale of the highest end AI
chips to China from Biden’s era face resistance – Nvidia wants to make money,
TSMC makes more money the more chips it makes, and China can/does retaliate
against other American companies who then lobby the US government against such
measures.
The US, under both
Biden and Trump, has put restrictions on which AI chips can be sold to
China. Unlike the USSR, to whom most chips were forbidden, a blanket ban is not
an option with China – China, unlike the USSR, is a huge market.
This then is the inherent contradiction at the heart of those AI chips – the US fears China becoming the AI leader, but the size of China’s market means the companies making those AI chips want to sell in China…
Comments
Post a Comment