(Editor’s Note: This post was submitted as a rebuttal to Andrew Chien’s July 24 SIGARCH Blog Post)
The recent post “Why Embodied Carbon is a poor Architecture Design metric, and Operational Carbon remains an important Problem” by Prof. Andrew Chien rightfully raises awareness of the challenges of reducing operational carbon. Specifically, we agree that one of the grand challenges of sustainable computing is matching electrical load with the availability of low-carbon electricity. However, the same blog post also controversially challenges the importance of embodied carbon.
In this post, we rebut the arguments raised against using embodied carbon as a design metric and conclude by advocating for more research on reducing embodied carbon. Unlike Prof. Chien, we assert that it is impractical and insufficient to rely on quickly deploying renewable energy to decarbonize manufacturing. Furthermore, the GHG Protocol, when used as intended, underscores the importance of embodied emissions and does not lead to “double counting”.
Throughout, we adopt the Greenhouse Gas Protocol scope definition as used in Prof. Chien’s post. From the perspective of datacenters, operational carbon includes Scope 1 direct emissions like diesel generators and Scope 2 indirect emissions from purchased energy. Embodied carbon is due to Scope 3, e.g., manufacturing IT hardware and datacenter construction.
Prof. Chien’s Arguments Against Embodied Carbon
Prof. Chien’s argument: Differences in reported embodied carbon are persistent and large across vendors and manufacturing processes
Rebuttal: Even with large variances from different data and models, evaluation of embodied carbon for different design choices is very much possible. The differences across vendors and processes are increasingly well tracked as the community is starting to build carbon quantification tools for embodied carbon, e.g., Greenchip, ACT and IMEC SSTS. These tools account for differences among fab locations and fabrication processes.
Additionally, large differences in a metric by themselves do not make a metric poor. Multiple research artifacts show the usefulness of embodied carbon (see conclusion). Additionally, our community already uses metrics that have similar orders of magnitudes in difference. For example, operational carbon emissions are highly time varying and even differ widely across regions.
Prof. Chien’s argument: Embodied carbon numbers are disproportionately large
Rebuttal: Large embodied carbon numbers are reported by companies who have set public goals to reach net zero emissions by 2030. Since abating emissions is costly, incentives are against companies inflating embodied carbon numbers. We thus believe that embodied carbon numbers are accurately represented. Additionally, companies like Meta and Microsoft are using extensive life-cycle assessments of their supply chains that are typically verified by independent third parties and experts across architecture, electrical engineering, supply chains, material science, and climate science.
Inattention to the embodied carbon metric is arguably part of the reason these numbers became large. Several recent studies have noted that trends like dark silicon have exacerbated the disparity between embodied carbon and operational carbon because energy-efficiency optimization has traded gains in the latter for increases in the former.
Prof. Chien’s argument: It costs only $1.4B to build renewable energy for TSMC’s 2.3GW capacity and TSMC’s customers are profitable enough to foot the bill
Rebuttal: This focus on renewable energy funding for TSMC is overly simplistic. Renewable energy sources and supporting infrastructures need to be designed, permitted and built. In Taiwan, specifically, there are land availability limitations and construction pauses during the typhoon season. Therefore, the $1.4B estimate vastly underestimates the costs of renewable energy.
The cost of recently-constructed wind energy in Taiwan is 7.7x higher per MW than assumed by Prof. Chien. Additionally, TSMC accounts for only 13% of wafer capacity and less than ~20% of the electricity consumption among the “Fab 5” (Samsung/Hynix/TSMC/Intel/Micron). There are also smaller vendors such as GF, UMC, and SMIC, along with a long tail of niche fab companies specializing in analog, PMIC, and RF components. This would require more than 5x the renewable energy capacity estimated by Prof. Chien.
Furthermore, wind farm capacity is typically stated as a peak power achieved at near-ideal operating conditions, and average power output for turbines in Taiwan is ~60-70% below peak. Unfortunately, fabs are not able to react to renewable energy production fluctuations – just like datacenters. Hence, significant additional investments into capacity and energy storage would be needed to match a fab’s daily load.
We also know that IC fabrication is only one part of the embodied carbon of computing hardware – according to Apple’s sustainability report, boards/flexes, aluminum/steel, displays, passive electronics, and assembly emissions in aggregate are comparable to IC production. Thus, we might assume that we need another 2x to decarbonize all of those vendors/suppliers.
Finally, electric power consumption of fabs makes up only part of their emissions. We thus conclude that decarbonizing the manufacturing supply chain will likely take decades and hundreds of billions of dollars.
Prof. Chien’s argument: Double counting of embodied carbon in the GHG Protocol means you must not sum or compare operational and embodied carbon numbers
Rebuttal: Prof. Chien’s blog post points to a paper by Bashir et al. whose argument we believe is flawed. Specifically, that paper discusses adding carbon emissions across companies. The GHG Protocol explicitly states that one should NOT sum the scope 3 emissions from different companies across supply chains. This misunderstanding leads the authors to claim that there is a “carbon multiplier” that overinflates embodied carbon, even for individual companies. This in turn led Prof. Chien and Bashir et. al. to make incorrect claims that embodied carbon is inflated and on a different “scale” than operational carbon.
The GHG Protocol does not have a double counting problem for the purpose it is defined for, namely carbon accounting by individual companies. As stated in the GHG Protocol, “scope 1, scope 2, and scope 3 are mutually exclusive for the reporting company, such that there is no double counting of emissions between the scopes within one company’s inventory”.
For any individual company that uses GHG reporting, sustainability reports account for actual emissions and the company’s scope 1/2/3 subcomponents can be added together. This accounting is frequently used in practice, e.g., to calculate carbon emissions for an internal carbon tax, such as Microsoft’s carbon tax.
Every gram of CO2e emitted into the atmosphere has the same effect. Thus, abating GHG emissions is an important challenge independent of when and where they are emitted. Increasing the use of renewable energy across the computing supply chain is a worthy goal and should be a focus of our policy and global stewardship philosophy. However, studies on IC manufacturing show that use of clean renewable electricity leaves substantial emissions for which there is currently no known solution. The same can be seen in top-down emission breakdowns which show that electricity makes up less than a quarter of global emissions. We need to make progress on both operational and embodied carbon emissions and that requires taking all scopes into account for architectural design decisions.
There are many open problems and research opportunities for reducing both embodied and operational carbon. In addition to the option of reducing chip sizes as mentioned in Chien’s blog post, there are many creative and constructive opportunities to reduce embodied carbon such as reusing entire HDDs or subassemblies like PCBs, reusing DRAM DIMMs, reusing other components, extending the lifetime of storage with limited endurance, reducing barriers to running servers for longer, using old devices in creative ways, batteryless mobile devices, and many others. Considering embodied carbon as a first-class optimization criterion, along with operational carbon, has the effect of reducing demand for carbon intensive products through techniques that span the computing stack, as well as propagating upstream the signal that providers should strive to also reduce their CO2 emissions.
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.