Why Curvy Design Now? Less Change Than You Think And Manufacturable Today

A curvilinear (curvy) chip, if magically made possible, would be smaller, faster, and use less power. Magic is no longer needed on the manufacturing side, as companies like Micron Technology are making photomasks with curvy shapes using state-of-the-art multi-beam mask writers today. Yet the entire chip-design infrastructure is based on the Manhattan assumption of 90-degree turns, even though in reality they are not manufacturable. In the past, transformative ideas such as the X Architecture have been developed but had limited adoption. So why curvy design now?

An expert panel during the 60th Design Automation Conference in 2023 addressed why the idea of curvy design is now a topic for research. The panel was moderated by Aki Fujimura, CEO of D2S, and combined expertise throughout the design chain with John Kibarian, CEO of PDF Solutions, Ezequiel Russell, Senior Director of Mask Technology at Micron, Andrew Kahng, Distinguished Professor at UCSD, and Steve Teig, CEO of Perceive. In the first blog about the panel discussion, I covered the perceived benefits of curvy design. In this blog, I’ll summarize the discussion on the real and perceived barriers to curvy design. The entire panel presentation is available here.

Aki Fujimura opened the panel with a critical observation that today’s GPU workstation has 340,000,000 times more computational capability than in the early days of Manhattan design, circa 1985. After the panel established the fact that today’s compute power is one factor enabling the photomask segment to adopt curvy masks today, Aki suggested it’s time to rethink curvy in EDA.

Less change in EDA than you may think

First, what we mean by curvy design refers to intra-connect (connecting transistors inside standard cells) and inter-connect (connecting standard cell inputs and outputs to other standard cell inputs and outputs) being curvy routes. Curvy routes will not necessarily be uniform width, but they are curvilinear metal paths for electrons to travel from outputs of transistors to inputs of transistors. An initial response might be to assume that the entire design tool chain must change for curvy design. However, Aki proposed that only four key areas need to change to enable curvy design, as shown in figure 1: custom design, routing, parasitic extraction and design-rule checking (DRC).

Fig. 1: Four things in EDA needed to enable curvy design. Source: Aki Fujimura, D2S.

Custom design today can handle curvy design shapes for applications such as photonics. But in Aki’s opinion, there’s an opportunity to work on custom design to make it more efficient for curvy design.

Automatic curvy routing will be essential, as will be parasitic extraction. Steve Teig finds the opportunity to reduce vias with curvy routing one of the most compelling reasons for the industry to develop and adopt it. Steve notes how Manhattan routing needs two vias at every turn whereas curvy routing rarely needs any vias for short connections in lower layers. Further, vias are physically unreliable so you must add design conservatism. In short, Steve calls vias “the enemy” (figure 2).

Fig. 2: Curvy routing rarely needs vias. Source: Steve Teig, Perceive.

DRC EDA tools will need to support curvy. Ezequiel Russell from Micron is a lithography expert for DRAM and flash memory with extensive experience with curvy photomasks for both 193i and EUV lithography. From his perspective, DRC – and for that matter mask rule checking (MRC) – become easier with curvy design. Ezequiel illustrates this for memory design in figure 3 indicating that Manhattan shapes produce countless DRC violations. In figure 4, Ezequiel illustrates how the curvy targets avoid DRC exceptions, are more lithography-friendly as they remove MRC constraints, and produce less deviation from optical proximity correction (OPC) mask to wafer shapes.

Fig. 3: Manhattan shapes produce countless DRC violations.

Fig. 4: Right side shows curvy targets avoid DRC exceptions, and curvy OPC is not constrained by MRCs and mask process capability. Source: Ezequiel Russell, Micron.

Curvy design is another potential scaling booster, but who will adopt it?

Andrew Kahng, Distinguished Professor at UCSD and Principal Investigator of the DARPA-sponsored OpenROAD Project, served as an advisor on the X Initiative and was a co-inventor of the Y Architecture, another idea leveraging non-Manhattan design for potential scaling, in 2005. Andrew characterizes curvy design as a new and mind-blowing paradigm for physical design. But his past experience prompts him to ask a number of questions, starting with which IC product organizations will be most excited about curvy design? You can view his detailed comments by downloading his slides here. Ezequiel quickly pointed out in the panel discussion that Micron is already targeting and using curvy shapes for masks. As an integrated device manufacturer (IDM), he sees there’s less friction for adopting curvy design within his model and believes that IDMs will be the first adopters of curvy design. Ezequiel also pointed out that High-NA EUV lithography is so expensive and complicated that the industry will want to use all the techniques possible, like curvy design, to extend current EUV and push out High-NA EUV as long as possible.

“I’ve seen this movie…”: Past attempts in the same genre haven’t taken root, so how will the curvy movie be different?

Andrew’s experience with X Architecture and Y Architecture leads him to question, “how will the curvy design movie be different?” He points out that since the early 2000s, there’s been more consolidation in every phase of the design chain. There’s competition for the attention and investment of R&D into next-generation technologies from back side power and 2.5/3D to row-based standard-cell architectures. John Kibarian of PDF Solutions points out that lithography scaling was working then, which probably limited the adoption of the X Architecture. It was easier to just scale to 65nm. But he contends that’s changed and lithography is no longer the solution. He further contends that curvy design and back side power will be synergistic. In fact, John thinks if you don’t do curvy design, you’re going to need to add more layers with the number of tracks in cell libraries reducing and interconnect dominating. Curvy design has the potential to reduce layers of interconnect and Ezequiel says that removing a layer for memory is a big deal.

Curvy design has the potential to be a transformative P&R technology

The final word in the panel went to Steve Teig, who summed up how curvy design could be transformative. First, wiring area could be reduced by up to 40% versus Manhattan design. Via counts could be reduced by ~50% versus Manhattan. Two layers of interconnect could be removed, or one layer removed in conjunction with reducing die size 10-15%. Steve’s final point is that curvy design could reduce performance conservatism due to manufacturing variability because curvy shapes are proven to be more reliably manufacturable. Sounds like the “curvy” design movie remake is likely to do better this time around.