For most of computing history, an architect optimizing a chip balanced two axes that everyone agreed on: performance and power. Both have mature, trusted modeling infrastructure built up over decades, which is precisely why design decisions can be argued quantitatively rather than by intuition. Carbon has had no such infrastructure. A June 2026 arXiv preprint, Architecture Carbon Tool v3: Enabling Sustainability-aware Silicon System Design Exploration, by Vincent T. Lee, Bilge Acun, Zachary Lewis and Carole-Jean Wu, argues that this is now a problem worth fixing, and presents the third generation of a tool aimed squarely at it.
The premise is that sustainability has moved from a corporate-reporting afterthought to an actual design metric. As the carbon cost of both manufacturing and operating semiconductor devices has come into sharper focus, the authors contend, an architect should be able to weigh carbon the way they already weigh joules and cycles — at design time, with a model, not after the fact in a sustainability report. That reframing is the heart of the paper.
"Like power and performance modeling tools, enabling sustainability-aware silicon systems design and optimizations will require a new generation of electronic design automation and architectural modeling tools."— arXiv:2606.16889 (Lee et al.), source
Why carbon is harder to model than power
The analogy to power and performance tools is apt but also reveals why carbon is the harder problem. Operating power is dominated by what the chip does while it runs; carbon splits into two very different components. There is embodied carbon — the emissions from manufacturing the device, which are front-loaded at fabrication and scale with die area, process node, and yield — and there is operational carbon, which depends on how much energy the chip consumes over its life and on the carbon intensity of the grid powering it. An architectural choice that lowers operational carbon, such as a larger cache or more accelerator area, can raise embodied carbon by consuming more silicon at an advanced node. Trading those two against each other is exactly the kind of decision a modeling tool needs to make tractable, and exactly the kind that intuition gets wrong.
ACT3 positions itself as an extensible and customizable modeling platform for research and advanced development, explicitly framed as paving the path toward sustainability-aware architectural design-space exploration. The emphasis on extensibility is a tell: carbon accounting depends heavily on assumptions — fab energy mix, transport, grid intensity, device lifetime — that vary by region and by year, so a credible tool has to let those assumptions be swapped rather than hard-coding a single number. A fixed carbon model would be obsolete the moment a grid decarbonizes or a fab moves.
What v3 adds over earlier ACT
Compared with previous versions of the Architecture Carbon Tool, the authors describe ACT3 as providing significantly richer modeling capabilities, enhanced collateral and analysis telemetry, and first-order design-space exploration capabilities. Each of those serves a distinct need. Richer modeling means the tool can represent more of the system rather than a single component in isolation. Enhanced telemetry means an architect can see why a configuration carries the carbon it does, decomposing a number into the contributions that drive it — which is what turns a score into an actionable design signal. And design-space exploration is the capability that elevates the tool from a calculator to a real optimization aid: instead of scoring one design after the fact, it can sweep across a space of designs and surface the carbon-aware trade-offs.
The authors are appropriately modest about the maturity of that last capability, calling it "first order." This is a technical brief rather than a finished optimizer, and it presents its case through several case studies illustrating ACT3's basic utility rather than a single headline benchmark. That framing matters: the contribution is the modeling infrastructure and the demonstration that carbon can be reasoned about during architectural exploration, not a claim that ACT3 has found the optimal low-carbon design for any particular workload.
Why this matters for the sector
The significance lands hardest in the part of the industry building large AI silicon, where both halves of the carbon equation are surging at once. Leading-edge nodes carry steep embodied carbon per wafer because of the energy intensity of advanced lithography and the long process flows, and AI accelerators run hot and continuously, piling on operational carbon over a deployment that can span years across data centers. An architect who can only see power and performance is, in effect, optimizing two-thirds of the problem and leaving the carbon consequences to be discovered later. A tool that brings carbon into the same design loop closes that gap.
The authors explicitly frame the release as an invitation, identifying opportunities where the research community can contribute to improving sustainability modeling and design methodology for silicon systems. That is the honest posture for an early-stage modeling effort: the value of a tool like this compounds with adoption and with shared, scrutinized assumptions about fab and grid carbon, and it is worth little if it stays a single group's internal spreadsheet. What the brief does not provide — and reasonably so, given its scope — is a validated comparison of its carbon estimates against measured fab data, or a standardized set of carbon-intensity inputs, both of which the community will need to agree on before carbon scores become as comparable across studies as power numbers are today.
The larger point stands regardless of those gaps. By insisting that sustainability is a design metric rather than a reporting obligation, and by shipping extensible infrastructure to act on that insistence, ACT3 nudges the field toward a world where a chip's carbon footprint is argued at the architecture review, on the same footing as its frequency and its power. For an industry whose products are simultaneously the most carbon-intensive things to manufacture and among the most energy-hungry to run, getting carbon into the design loop early is not a soft concern — it is where the largest, least-reversible decisions already get made.