Offshore

1 view19:02

Antonio Linares
$AMD is poised to challenge $NVDA's stronghold in the AI sector for two key reasons:

1. Traditionally, $NVDA has maintained a competitive edge through CUDA, a platform that ensures smooth interactions between developers and $NVDA GPUs.

Yet, the dynamic has shifted with the rise of Pytorch as the leading framework for deep learning (AI), now seamlessly integrating with $AMD GPUs right out of the box.

$AMD has now recently opened up ROCm (its CUDA equivalent) to the open source community.

This integration marks the beginning of the erosion of $NVDA's software stronghold, leveling the competitive landscape.

2. In contrast to $NVDA’s uniform chip design, $AMD employs a chiplet-based architecture.

This progressive strategy has already overturned $INTC's market dominance by offering high-performance chips at reduced costs.

The chiplet design has a unique benefit: if one component fails, it doesn’t compromise the entire chip.

As a result, chiplet architectures typically achieve much higher yields than their monolithic counterparts.

However, adopting chiplet technology implies a shift toward thinner margins, which could negatively affect $NVDA's financial health.

This represents a classic instance of the Innovator’s Dilemma – the hesitance of entrenched firms to self-disrupt.

$NVDA has technically already pivoted to chiplets with the launch of the Blackwell architecture, but there are some important considerations.

A thorough analysis of Blackwell's architecture indicates that the chip fundamentally comprises two large chips that are interconnected.

Previously, $NVDA architectures followed a similar pattern.

However, Blackwell marks a new development where for the first time, its two chips function cohesively at both the software and network levels.

Essentially, Blackwell consists of two chiplets, signaling a cautious initial move towards a chiplet-based architecture.

Each of these chiplets, however, are already at the boundaries of the maximum reticle size.

As a result, $NVDA is nearing the physical limitations that make the production of monolithic chips increasingly difficult as we progress to smaller process nodes.

To increase computational capabilities, $NVDA now faces the challenge of either dramatically escalating the complexity within each of the two existing dies or adding more monolithic chips to the Blackwell architecture.

Should $NVDA not completely shift towards $AMD's strategy, there are two potential long-term outcomes:

1. $NVDA maintains its lead by essentially linking more monolithic chips, each at the reticle size limit, akin to a pseudo-chiplet strategy, with $AMD remaining a peripheral contender.

2. This approach may not prove scalable compared to $AMD’s dedicated chiplet architecture, potentially allowing $AMD to capture significant market share at the high end of the spectrum.

What's clear at this stage, however, is that $AMD has a structural advantage to bring AI compute engines with a differentiated price to performance ratio to the market.

Time will tell whether $AMD can prosper on the software side, which is essential for its hardware endeavors to bear fruit. But $AMD is taking the right steps to get there.
tweet

1 view19:02