Who Will Regulate Data Exchanges In Chiplets?

Scaling is still important when it comes to logic and low power, but it’s no longer the main avenue for improving performance. What used to be a single chip, comprised of various IP blocks and components on a single SoC, is giving way to a heterogeneous collection of chiplets — at least for the big chipmakers and system companies at the leading edge.

Chiplets are currently the best solution for improving performance per watt for an entire system, which may include 3/2/1.x nm logic, as well as a 90 nm analog block and a collection of different memory types. The chip industry has been racing to figure out how to build these decomposed SoCs in a way that is both cost-effective and customizable for different markets, from data centers to automobiles.

In fact, the upcoming Hot Chips conference is expected to showcase heterogeneous architectures, integrated together using type of advanced package, and tied together with some common interconnect, such as the Universal Chiplet Interconnect Express (UCIe). Hot Chips is typically where companies like IBM, Intel, AMD, NVIDIA, Google, and even Tesla show off their latest architectures. This year’s lineup includes everything from discussions about UCIe to memory-centric approaches.

Chiplets so far have been almost entirely internally sourced, unless you consider HBM a chiplet. (Some say it is, more say it isn’t, at least for now.) But as the market shifts to third-party chiplets, it will force other changes as companies seek to integrate different chiplets into their designs.

The chiplet concept is similar to LEGOs, providing you consider effects such as noisy neighbors, bonding, hot spots, and different use cases. These designs also need to factor in how large a role AI/ML/DL will play, because those computations can get very hot for extended periods of time. This is why commercial chiplets initially may look more like IP sub-systems than IP blocks, where much of the integration is done already and the characterization is well-documented.

All of this opens the door to another challenge, though, which has been given scant attention. For the commercial chiplet marketplace to work effectively, companies will need access to data from other vendors, and they will need to figure out a way to share it without giving away internal secrets. This is harder than it sounds, because permissions and authentication methods need to be crafted and in-place at the architectural level, accessed throughout the design-through-manufacturing flow, tracked for as long as that data holds value (unless there’s a one-time purchase price) and verified at multiple points that are still yet to be defined.

It’s not at all clear what constitutes fair use of data, which is a company’s IP. And this becomes even more complicated once this data seeps out into the public domain through ChatGPT or other generative AI platforms that are still being developed. How will data be regulated? And how will it be perceived in case law that doesn’t exist yet. So far, there are no answers.

The chip industry is heavily focused on just getting these devices to function. But the really hard part may still be ahead — business discussions, legal contracts, and lots of discussions about how to trade sensitive data without losing it to a potential competitor.

Related 
Chiplets Knowledge Center
Index to chiplets articles
Chiplets: Deep Dive Into Designing, Manufacturing, And Testing
EBook: Chiplets may be the semiconductor industry’s hardest challenge yet, but they are the best path forward.