First, what is an architectural metric? A quantitative measure that can be used to inform computer architecture choices to compare choices such as organization, mechanisms, and implementation [ABB64,HP94]. The best metrics would be applied to measure specific architectural designs, and thereby support architectural design insights that inform future hardware architecture. My objective was to catalyze a technical discussion about architectural metrics for sustainability and how we might improve them! The two key arguments in [ECP] that EC is a poor architectural metric are
EC varies widely by vendor and process; its defined by location and company used to manufacture. The two examples of logic fab and Hydrogen are chosen as extreme examples. But it appears likely we will have nearly identical products with radically different embodied carbon soon. First, Nvidia is publicly exploring sourcing from Intel Fabs for GPUs [Huang23], perhaps producing perhaps identical Nvidia GPU’s with 8x different embodied carbon due to vendor of manufacture. If that doesn’t happen, TSMC (Arizona) and Samsung (Texas) are building multiple fabs (TSMC, Samsung) that will produce similar outcomes for differences in embodied carbon. In both cases will make architectural conclusions based on embodied carbon difficult.
The analysis of decarbonization costs was to highlight that the high value and profitability of computing hardware (eg Apple, Nvidia, AMD, Intel, etc.) and computing services (eg. Microsoft, Google, Amazon, Meta, etc.), creates a large profit pool, easily large enough to pay for decarbonization with current technology. The profitability is important, as in many other areas where carbon emissions reduction is essential such as global agriculture [g1,g2] there is scarce profit pool to fund decarbonization. We should be optimistic that computing can and will drive rapid decarbonization of its supply chain, an optimism shared by others such as Apple who are driving ambitious plans for 100% power decarbonization in their supply chains by 2030 (including commitments by suppliers such as TSMC, SK Hynix, ST Micro) [Apple23].
Second, what might be better architectural metrics for sustainability? This is a difficult question, and a topic on which research is needed. Good metric properties include increasing strongly with the e-waste and energy use associated with hardware manufacture. Other good properties would be stability over upstream supply chain changes (eg. more renewables, or a new low-power lithography method, low water-use technology for wafer washing, or a new catalyst for a key chemical process). Example candidates “silicon area, in a X nm process”, or “normalized silicon area”.
Finally, one might ask “other fields such as building design use embodied carbon, what is different here?” The challenge here is the distinction between an architectural design/implementation, and a specific product (sku). We might consider two designs for computing (Nvidia H100, Intel Sapphire Rapids 8470), and compare their performance on ML benchmarks (mlperf) and their silicon cost (logic area, memory area) to assess cost-performance. However, if we substitute embodied carbon for silicon cost, by the argument above, we find that embodied carbon may depend little or not at all on the architectural design – rather just the country or fab that the chip is manufactured. This product (sku) is the determinant of embodied carbon (and location of manufacture), overpowering the architectural comparison. However, populating the datacenter (akin to building design) is about selection amongst available products (sku’s). So the the nuance here is that embodied carbon may be useful as a product/sku metric, but not good as a computer architecture metric.
The authors of [RB] argue that EC can be used to advance useful architectural research, citing diverse approaches to extend lifetimes. I applaud these approaches, and the IOT/pervasive computing community has long explored approaches such as battery-less [Telos,WISP,MMos, BL] and lifetime extension to reduce environmental damage from computing [Pervasive23Panel], so other metrics can be useful to drive manufacturing sustainability.
Let’s have a constructive, open, scientific debate to find good architectural metrics for sustainability!
About the author: Andrew A. Chien is the William Eckhardt Distinguished Service Professor at the University of Chicago. His research interests include parallel computer architecture, cloud software, datacenters, programming systems. He is the leader of the Zero-Carbon Cloud project, and a leader for computing sustainability. From 2017-22, Dr. Chien served as Editor in Chief for Communications of the ACM, and from 2005-2010 as Vice President of Research at Intel. He has held Professorships at UCSD and Illinois (UIUC), and is a Fellow of the ACM, IEEE, and AAAS. Dr. Chien received his BS, MS, and PhD from the Massachusetts Institute of Technology.
Disclaimer:These posts are written by individual contributors to share their thoughts on the Computer Architecture Today blog for the benefit of the community. Any views or opinions represented in this blog are personal, belong solely to the blog author and do not represent those of ACM SIGARCH or its parent organization, ACM.