The Chip Insider®– The Nvidia versus AMD AI Chiplet Battle

Author: G. Dan Hutcheson

 

  5 Min Read     September 16, 2025

 
 

The Chip Insider®– The Nvidia versus AMD AI Chiplet Battle

The Nvidia versus AMD AI Chiplet Battle: Why does Nvidia still have chips using the reticle size limit if it can save costs by packaging smaller dies? The simple answer is that it’s because Nvidia’s design requires that many die. Nvidia is not optimizing for cost. Nvidia is on a technology strategy that optimizes for maximum performance to deliver higher customer value… AMD’s on a cost strategy…

The collision of AI and lithographic limits created the imperative to break away from monolithic ICs and move towards chiplet strategies... GPU architectures made it easy to break workloads up and process them in parallel… This meant that more transistors were always better if the AI chip was architected correctly. There were no demand-side limits. Only supply-side limits from manufacturing.

So, an Nvidia B200 chip has 208B transistors, packed into two chiplets. AMD’s newest MI350 has 185B transistors… with 8 … chiplets... AMD’s value proposition is that it’s the lower-cost alternative to Nvidia. If Nvidia is a Ferrari, AMD is a Porsche. The key tradeoff to understand here is that adding die-to-die connections via chiplets has a … performance/W loss when disaggregating a design into more chiplets. But it does lower cost by increasing die yield…

Now, if the strategy is a technology one, why not pack all the transistors into a single monolithic die? … ASML’s 0.55 Hi-NA EUV tool cuts the maximum field size in half from… History doesn’t always repeat itself. But AMD may have the performance advantage due to its far greater experience with chiplets. But it still doesn’t solve the need for the AI community to solve its power consumption threat. For this, the lithography industry is working on a 6-by-12-inch reticle, which would bring back the field size. It’s retrofittable to existing Hi-NA EUV tools, but won’t be ready until 2030. Either way, a critical competitive difference in AI chips between AI and AMD with regard to chiplet choices will be in their lithography tool choices.

Maxims: There are only five business strategies: cost, quality, distribution, technology, and intellectual property (IP).

"History never repeats itself but it often rhymes.” – as Mark Twain is reputed to have said

 

Access the Latest Edition of the Chip Insider

Stay ahead of evolving trends with data-driven insights and expert analyses of the semiconductor industry.

 

TechInsights

 
LinkedIn
X
YouTube
App Store
Google Play Store
 
 
EcoVadis
ISO 27001 Certified