← Back to Library

Why Amd's chiplets work

Most industry analysis fixates on the race for smaller transistors, but Asianometry makes a compelling case that the real revolution in computing isn't about shrinking silicon—it's about breaking it apart. The piece argues that the era of cramming more components onto a single chip has hit an economic wall, forcing a pivot to a modular architecture that was once considered a niche failure. This is not just a technical shift; it is a fundamental rethinking of how value is created in the semiconductor supply chain.

The Economics of Breaking Up

Asianometry begins by dismantling the assumption that "bigger is better" in chip design. For decades, the industry relied on Moore's Law, but the author notes that "the increasingly difficult economics of designing and fabbing a faster chip has opened the door to new approaches." The core problem is yield: as chips grow larger to accommodate more transistors, the probability of a fatal manufacturing defect rises exponentially. "Doubling the number of transistors on a single chip more than doubles the cost," Asianometry writes, highlighting a brutal mathematical reality that threatens the viability of monolithic designs.

Why Amd's chiplets work

This framing is crucial because it shifts the conversation from pure physics to hard-nosed business logic. The author points out that while mobile phone chips can absorb the massive costs of leading-edge nodes due to sheer volume, server and graphics processors cannot. The solution, as the piece explains, is to abandon the "monolithic" approach in favor of "heterogeneous integration." Asianometry writes, "It may prove to be more economical to build large systems out of smaller functions which are separately packaged and interconnected." This quote, originally from Gordon Moore but revitalized here, serves as the historical anchor for a modern strategy.

Critics might note that this approach introduces new complexities in thermal management and signal latency, which the author acknowledges but treats as solvable engineering hurdles rather than dealbreakers.

The Interconnect Problem and the AMD Solution

The commentary then pivots to why this strategy failed in the 1980s and 90s but succeeded now. The author argues that previous attempts at multi-chip modules failed because "vendors never really could wrangle the higher level performance problems of integrating multiple different dies solely through packaging." The bottleneck was the interconnect—the bridge between chips.

Asianometry identifies the turning point as the development of the Infinity Fabric by AMD, a proprietary interface that managed data transmission with high bandwidth. "The system is an evolution of something that AMD has used before for connecting items on a motherboard via sockets," the author explains. This innovation allowed AMD to bypass the yield issues of massive chips by splitting them into smaller, manageable pieces. For their first-generation server chips, codenamed Naples, this meant "saving up to 40 percent on costs" compared to a monolithic design.

"Splitting that into chiplets vastly lowers the cost of the chip assuming performance remains roughly the same."

The author's analysis of AMD's "Naples" and "Rome" processors illustrates the power of this modularity. By using a leading-edge 7-nanometer process only for the critical computing cores and a cheaper, legacy 14-nanometer process for the input/output functions, AMD achieved a "drastic enough savings to overcome extra costs on the packaging side." This is a masterclass in cost optimization, proving that not every part of a chip needs to be built with the most expensive technology.

Strategic Implications and Market Scale

The piece concludes by examining why this strategy works for AMD specifically and whether it can be universally applied. Asianometry argues that chiplets are "not a panacea" but rather a response to a specific set of circumstances. The author writes, "AMD just found itself in the right time and place for it," citing the company's ability to design leading-edge chips and its presence in massive markets like gaming and data centers.

The modularity allows for a flexibility that is impossible with monolithic designs. "Now AMD can wait and review the market conditions before deciding whether to create more 24, 32, or 64 core epic server products," the author notes. This ability to mix and match chiplets across different product lines—from desktops to servers—creates a scale that helps amortize the astronomical costs of modern fabrication.

However, the argument leaves some questions about the long-term dominance of this model. As competitors like Intel and TSMC develop their own packaging standards, the proprietary advantage AMD enjoys with Infinity Fabric could erode. The author hints at this, noting that "big companies like Intel and TSMC are starting to offer their own packaging IPs and standards into the market space," suggesting that the future may belong to those who can standardize the interconnects, not just those who design the chips.

Bottom Line

Asianometry's strongest contribution is reframing the chiplet strategy not as a technical novelty, but as an economic necessity born from the limits of Moore's Law. The piece effectively demonstrates that the future of high-performance computing lies in modular design, where the value is created through intelligent integration rather than sheer transistor density. The biggest vulnerability in this argument is the assumption that the interconnect bottleneck can be permanently solved; if data transfer speeds between chiplets lag behind processing speeds, the entire architecture could face diminishing returns. Readers should watch how the industry standardizes these connections, as that will determine whether chiplets become the universal standard or just another niche solution.

Sources

Why Amd's chiplets work

by Asianometry · Asianometry · Watch video

for the first six decades of the semiconductor industry manufacturers improved performance by cramming more transistors onto a single die it wasn't easy but it worked but now the pathway ead isn't so clear the increasingly difficult economics of designing and fabbing a faster chip has opened the door to new approaches for achieving better performance one of the more popular approaches is what amd is doing with chiplets in this video we dive into how it works and why it now makes sense in this new semiconductor world but first let's talk about the asian armature patreon if you like what this channel does you can support the work by joining the early access tier early access members get to see new videos and selected references for them before they are released to the public it's not a lot of money and i appreciate the support thanks and on with the show today it costs more than ever to design and fab a chip with a leading edge process node to scale transistor density every couple years or so seemingly magical solutions like euv lithography cost billions to develop and hundreds of millions to deploy the customer has to make the economics work on that one which means that nodes like five nanometers and the such can only be deployed at a titanic scale for instance like with mobile phone chips that sell in the millions each year without that scale and without access to greater transistor densities how then to deliver increasingly better performance one thing that gpu and server cpu companies have tried is to make the chip bigger chips mean more transistors despite lower densities however total soc size has a ceiling imposed onto it by lithographic mask sizes generally about 858 square millimeters in other words it's impractical to try to etch a chip design that's too large furthermore large socs have yield concerns bigger chips have a bigger chance of having rors on them an unrepairable manufacturing fault can ruin the whole soc a loss that the chip seller not the foundry has to eat if it can't find some way to use them elsewhere such yield costs do not scale linearly in other words doubling the number of transistors on a single chip more than doubles the cost thus the industry realized that it had to find new ways to deliver better performance ...