Design-Space Exploration for Macro-Based SRAM Configuration on the SKY130 PDK
This paper introduces an automated design-space exploration framework for SKY130 macro-based SRAM configuration that evaluates nine memory-type and placement combinations using realistic area and leakage metrics, demonstrating that the choice of candidate selection method—specifically a Pareto-optimality filter versus traditional heuristics—significantly impacts the final design outcomes.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Inside the tiny silicon heart of modern electronics, a specific kind of memory called SRAM acts as a high-speed workspace for processors. Designers building these chips often face a frustrating mismatch: the tools provided by the chip manufacturer offer only a few fixed sizes of this memory, like standard bricks in a box, but the actual project requires a custom shape or a specific amount of storage. For years, engineers have had to manually piece together these fixed bricks to fit their unique needs, a process that is slow, prone to error, and leaves no room to check if a better arrangement exists. This is the problem researchers at the National Institute of Technical Teachers' Training and Research in Chennai set out to solve, working with a specific, open-source manufacturing kit known as SKY130.
The team focused on a tool called the Macro Memory Cell Generator, which already existed to help stitch those fixed memory bricks into custom sizes. However, that original tool acted like a passive assembler: it would take whatever memory type and layout style an engineer typed in and build it, without ever asking if a different combination might be smaller, more energy-efficient, or easier to wire up. The researchers realized that simply building the first thing that fits was not enough. They needed a way to automatically explore all the possible ways to arrange these memory blocks and then pick the absolute best one based on real-world constraints.
To do this, the team built a new search engine that runs alongside the existing tool. Instead of guessing, the system generates every possible valid arrangement for a given memory size. It then measures three critical things for each arrangement: how much physical space it takes up on the chip, how much electricity it leaks when idle, and an estimate of how long the wires connecting the blocks would need to be. Crucially, the researchers did not rely on simple math formulas to guess the space required. They found that formulas often miss the extra room needed for wiring, so they forced the system to run the actual placement software used in chip design to get a precise measurement of the area. This ensured the results reflected reality, not just an idealized theory.
The researchers then had to decide how to pick the winner from the list of candidates. A simple rule like "pick the smallest one" ignores the other important factors like power and wiring. To solve this, they used a method called Pareto-optimal selection. Imagine a group of runners where you want the fastest, the lightest, and the most experienced. A runner is only considered a top contender if no other runner beats them on all three traits at once. If one runner is faster but heavier, and another is lighter but slower, both stay in the running. The system filters out any option that is clearly worse than another in every way, leaving only the truly competitive choices. When more than one option remains, the system uses a balanced score to break the tie, weighing space, power, and wiring equally.
To ensure this approach was actually better than other ways of choosing, the team tested their method against five other established decision-making techniques. They ran the same set of fifteen different memory configurations through all six methods. The results were surprising: the methods agreed on the best choice in only about one-quarter of the cases. In the other three-quarters, the choice of method completely changed the outcome. Three of the methods, which used very different logic, always agreed with each other, but the team's Pareto-based approach often picked a different, valid option that the others rejected. This proved that the way a designer chooses a configuration is not a minor detail; it fundamentally changes the final chip design.
The researchers also added a feature for designers working with a strict physical limit, such as a fixed block of space on a chip that cannot be exceeded. The new system can take a maximum width and height as an input and automatically discard any memory arrangement that is too big, returning only the best option that actually fits inside those walls. In one test, a configuration that was the smallest overall was rejected because it was just slightly too tall for the allowed space, and the system correctly fell back to the next best option that fit. This mimics the real-world scenario where a designer has a specific hole to fill and needs the best memory that fits inside it, rather than the best memory in the abstract.
The study concludes that while the original tool could build custom memory, it lacked the intelligence to choose the right configuration. By automating the search, using real placement data instead of estimates, and applying a rigorous method to select the best candidate, the researchers have provided a way to make these custom memories more efficient. They found that relying on a single, simple rule for selection is risky, as different methods lead to different results. The work remains a simulation-based exploration, meaning the wire length numbers are estimates that have not yet been validated against actual post-route wirelength; consequently, the configurations identified in the study should be read as provisional until that validation runs. The framework is now open for others to use and test against actual chip manufacturing flows. The key takeaway is that in the complex world of chip design, the choice of how to pick a memory layout is as important as the layout itself, and leaving that choice to chance or a single metric is no longer necessary.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.