Chiplets: A Short History | EE Times
In another “rabbit hole moment,” I lost myself in a bout of reminiscence on the Internet Archive’s Wayback Machine. A prehistoric SemiSerious blog post from 2009 attempted to provide updates from ISSCC 2009. TechInsights engineers Aaron Murray, James Bull, and Mohammad Ahmad provided the boots on the ground reports to HQ. I believe these three fine gentlemen all continue their yeoman’s work at the Ottawa company.
At ISSCC 2009, Intel was presenting a number of papers related to its Nehalem processor. Mark Bohr’s plenary indicated that Intel was putting a lot of effort into system-on-chip (SoC) development. The Core i7 Nehalem design pulled the memory controllers and DDR3 I/O’s onto a common substrate with the CPU cores. At the time, Intel suggested that they were looking beyond the traditional PC market and into hotter ones like mobile. In Bohr’s words, “Intel is no longer a one-size-fits-all company.”
From this, there was a recollection of a “great debate” between SoC and system-in-package (SiP) advocates. Intel had been focused on package integration while Texas Instruments pounded the SoC drum. Now it would be reasonable to ask what relevance these two companies might have on the 2021 landscape considering that other players are trending higher in the conversation than these two old stalwarts. Gordon Moore’s company and the birthplace of his law was already thinking about More than Moore (MTM) back then.
The idea of the disagreggation of integrated circuit design into smaller physical pieces of silicon, chiplets, that can be mixed and matched Lego-style at the package level has been generating buzz for some time. For Intel’s part, they have not abandoned the SoC mentality but continue to dispel the one-size-fits-all label. From Embedded Multi-die Interconnect Bridge (EMIB) to their central role in the State-of-the-art Heterogeneous Integration Prototype (SHIP) program, it appears Intel will not be left behind. Integration of diverse designs using more specialized chiplets fit Intel as much as any other company whether direct competitors like AMD with their EPYC implementations or vendors with completely different market aspirations.
In the 2009 ISSCC coverage, I predicted that Intel would be ready for mobile based on this renewed focus on the SoC. Looking back, it’s easy to see how wrong I was. Fast forward, and the big data age and AI accelerators are much more interesting to a company like Intel. Acquiring an FPGA company (Altera) in 2015 helped maintain Intel’s territory in the data center as cloud computing drove a more rapid advance in technology as well as market importance.
Volatile vernacular
There have been a few incarnations of terms along the way. Anyone who has posted on topics like heterogeneous integration can relate a story or two of commenters reminding them that only the term, not the technology, is new. But cut some slack here. We are all just trying to fill column space.
In general, heterogeneous integration was a term adopted to sell MTM as an approach to keep innovation moving and industry economics headed in the right direction. A seminar or conference or at least a conference call between marketing departments would have helped. The original vernacular was heterogeneous integration technology (HIT). ASE Group truncated this to HI. Hello right back at you.
IEEE went so far as to create a new pronoun with its heterogeneous integration roadmap (HIR).
In a land before time inhabited by engineers who may tell you fond stories of rubylith days, there were hybrid integrated circuits. Hybrids were collections of different component types to create a functional module. The hybrids were definitely heterogeneous. In that sense, today’s lingo is just re-branding with the help of a thesaurus.
Moving on, there was the multi-chip module, or MCM. These were perhaps less heterogeneous than the hybrids since multiple die were assembled onto a common package substrate. But several functions and even multiple material system ICs, for example silicon along with GaAs, were collected into these platforms. Adding a discrete or two, maybe some decoupling capacitors, would definitely make this another hybrid or heterogeneous application.
Someone who was more intimately involved at the time could help shed light on this, but it would seem to me that MCM was meant to denote a change where multiple individual integrated circuit die appeared on a common substrate. That was the milestone that spawned the new term.
The MCM is still important today, decades after the technology was introduced. With the advent of newer enabling technologies like through-silicon vias (TSVs) and 2.5D integration with silicon interposers, a More than Moore strategy moves the industry toward the 3DIC. The most influential products in this field are the so-called high bandwidth memory (HBM) products. One of these seemed to acknowledge history rather than inventing a new term. The Hybrid Memory Cube (HBC) product co-developed by Micron and Samsung used the latest TSV approach to stacking multiple DRAM die with their controller.
One term has more or less persisted. System-in-package (SiP) has been around for a couple of decades. That’s an eternity considering the pace of change, in terminology at least.
The ASE Group introduction to HI illustrates and explains collecting a set of dissimilar die into a single SiP. This is an obvious representation of the chiplet concept, but the term does not appear on the page.
Chiplets rising
One should never discount the importance of terminology and marketing to technology. There is spin, re-spin, regurgitation and recycling.
Enter chiplets.
Why not? It’s easy to say, and easy to digest the concept. It’s much better than the nerdier multi-chip module (easy to interpret) or heterogeneous integration (open to interpretation).
To get an idea of when the term became popular, Google Trends provides a history of chiplet as a search term going back to 2004. There was an early dalliance with the chiplet, but that died down quickly. The real uptick has been since 2019.
Google Trends is useful, but it will not reveal the context in which chiplet was searched. Rather than stand by the assumption that the chiplet we are considering took off only two years ago, a look at patent activity will be more revealing. This information trove gives us a much longer timeline as well.
A search of the patent databases reveals use of the chiplet term as early as 1969. However, in the integrated circuit field, it is only in IBM applications published in late 2000 that our current understanding of the term and technology align.
Although, this early use of chiplet is not in the heterogeneous sense we understand to mean a SiP of heterogeneous die. The early days of chiplets were related to driver chips for LCDs. And that makes sense if you have looked closely at any display panel larger than a cell phone. Whether on flex substrates attached to the panel glass or directly on the panel, there are many tall and skinny LCD column driver chiplets attached to each large LCD panel.
Looking at this from a trending term perspective, it is fair to say that use in a reasonable volume of patent applications would provide an indication of the adoption of the term. In that sense, chiplets really pick up the pace in 2010. It is worth noting that several entities and inventors involved in display technology continue to represent a a large proportion of the activity related to chiplets. However for system-in-package heterogeneous integration, IBM, GlobalFoundries and Intel are the most active companies.
Like it or not, “chiplet” is here to stay. “Googling” it and patenting it make it so.
It seems like going down a rabbit hole loves company. It’s time we all came up for air.