Chapter 7 Digital Humanities and the Energetics of Big Data

The Energetics of Data Production and Preservation

To understand how the energy consumption of big data stacks up against that of nondigital media, we build upon the insights of Gilbert Shapiro, John Markoff, and Silvio R. Duncan Baretta (115–17). Their framework, which assesses the likelihood of a historical document’s survival, encompasses various phases: recording, reproduction, preservation, cataloging, and publication.4 Our back-of-the-envelope calculations estimate the energy costs associated with each juncture, factoring in both the embodied energy inherent in the materials used for crafting these records and the cumulative human and machine labor for their reproduction and preservation (see Table 7.1).5

When considering the energy overheads of nondigital and digital media, a dichotomy emerges. The creation and replication of nondigital artifacts, while more energy-intensive initially, starkly contrast with that of digital materials, which demand consistent energy input for their sustenance and access. The history of information media uncovers a trend toward the diminishing of energy needed to record a given amount of content. The process of inscribing text onto bronze, such as those found in early China’s ritual vessels, required about 3 GJ, and an equal amount of energy was needed for its replication. A substantial part of this energy was invested in the creation of the bronze itself. In contrast, transcribing manuscripts onto paper or parchment drastically slashed this energy demand to between 0.5 to 2.5 MJ per page, even though the same amount of energy was still needed to make copies of the original.

The introduction of print marked a leap in reducing energy usage, particularly for reproduction. While engraving a page onto a woodblock demanded about six to sixty times more energy than manuscript writing, the subsequent printing process drastically cut the energy costs to about 200 kJ per page, a mere tenth of the energy needed for manual transcription. Advancements in printing technology continued to drive down these costs. The energy required for producing the first copy of a page dropped to about 2 MJ for hand letterpress and under 1 MJ for rotary print. Digital technology has decreased energy consumption, eclipsing all prior media forms in volume and reach. Digital text production outpaces the finest print technologies by a factor of 10 in energy efficiency, and digital replication surpasses a two-thousandfold increase in efficiency.

The remarkably low production and reproduction costs of digital media come with caveats: reduced longevity and higher upkeep demands. Volatile memories, such as L1, L2, and L3 caches and dynamic random access memory (DRAM), immediately lose the loaded data in the event of power interruption. Nonvolatile storage, including HDDs, SSDs, optical discs, and magnetic tape drives, require electricity, albeit for different reasons. To avert data deterioration, these devices require climate control, with optimal conditions for magnetic tapes and optical discs being the temperature range of 15–25°C and humidity levels between 30 percent and 50 percent (CLIR). HDDs, particularly helium-sealed drives, demonstrate greater resilience and tolerate a broader temperature range, from −40–70°C (AKCP), and they need to be shielded from direct sunlight. In addition, consider how modern cloud computing categorizes data as hot, warm, or cold according to usage patterns and energy use. SSDs designed for handling high-demand or “hot” data, averages between seven and ten years, while SSDs without access to a power source may risk data loss over time due to the degradation of their NAND flash components.6 A cold-storage facility like Meta’s Prineville location stores infrequently accessed data on low-energy archival-grade HDDs and Blu-ray discs, with shelf lives of thirty and one hundred years, respectively (Bandaru and Patiejunas; Miller; Hogan). Millenniata’s M-DISC and Microsoft’s Project Silica exceptionally provide data preservation on media designed to last more than 1,000 years, but their adoption remains limited (see Hachman; Salter). In the ICT sector, linear tape open (LTO) tape drives, with about thirty years of service life, are a favored backup solution (Hewlett Packard Enterprise). Most of these digital memory and storage systems rely on continuous power for data longevity.

In contrast, preserving paper primarily involves maintaining a dry environment and avoiding temperature fluctuations, while bronze requires minimal upkeep (see Figure 7.1). The lifespan of digital media, usually ranging several years to a couple of decades, is significantly less than the centuries that paper can last or the millennia for bronze artifacts. The ICT industry’s strategies for preserving big data via redundancy and duplication mirrors the way that historical records have been preserved by repeatedly creating copies. For example, updates to early modern Chinese genealogies occurred approximately every sixty years (Pieke, 113), while reproductions of classical texts took place about once per century. One could argue that digital media’s need for frequent duplication diverges from the slower duplication processes seen with other media. For our purposes, the aggregate energy needed to preserve and maintain the integrity of big data in cloud facilities is the ultimate dilemma. Even though each act of digital replication consumes negligible amount of power, the aggregate effect of these repeated operations leads to substantial energy use over time.

To sum up, the power needed for digital publication is considerably less than that for a printed book (which in turn is more energy-efficient than manuscript production), and far less than the energy needed for creating epigraphic materials. Media with higher energy demand at the moment of creation typically boast longer lifespans, while those requiring less energy facilitate a greater volume of work. This dynamic results in digital materials outnumbering printed media, with printed works surpassing both manuscripts and epigraphic artifacts in quantity. Scholars specializing in early periods tend to depend on materials preserved on more enduring media or those that have been replicated over time, typically missing out on ephemera. In contrast, researchers focused on more contemporary periods have access to a plethora of ephemera, but face uncertainties regarding their long-term preservation.

Paradoxes and Trade-offs

The ways that a modern DH researcher’s work is dependent on Big Tech infrastructure come at the cost of energy consumption and environmental footprints, serving as yet another reminder of the paramount importance of recognizing the materiality of digital technologies. To reference a few well-known relevant studies on this topic, Friedrich Kittler’s “There Is No Software” has significantly contributed to our understanding of the physical foundations of digital technologies. Matthew Kirschenbaum’s Mechanisms and Bitstreams have demonstrated the application of forensics in the study of electronic literature and digital collections. Kate Crawford’s Atlas of AI, in turn, has taken a critical look at the AI industry’s dependence on precious minerals, water, and electricity (23–51), extending beyond Kittler’s perspective, which was limited to silicon. Thomas Mullaney’s introduction to Your Computer Is on Fire also stresses that “nothing is virtual” (5). Further, Nathan Ensmenger urges a reconsideration of “the Cloud as a factory, and not as a disembodied computational device” (43). Mullaney’s advocacy for a critical stance toward the tech industry, described as a “call to arms” with “unapologetically direct and bold arguments” (8), and Crawford’s examination of the AI industry merit careful contemplation. At the same time, it is also important to be specific and precise. Contrary to the centrality of lithium in Crawford’s discussion of AI, it is a critical mineral used in the production of batteries for mobile devices and electric vehicles, not the semiconductors, data storage units, and telecommunication networks that underpin cloud computing and AI.

Adding to this challenge is the Jevons paradox, which observes how efficiency gains encourage more utilization. “According to a principle recognized in many parallel instances,” William S. Jevons astutely noted, “new modes of economy will generally lead to an increase in consumption” (103). Originally articulated in the context of coal use, the Jevons paradox offers a compelling lens for understanding the burgeoning of Web 2.0 services, in which massive capital investments and the development of centralized infrastructures drove efficiency gains. The technological strides in recent years have resulted in the seamless weaving of digital technologies into various facets of daily life, including academic research.

Recent market data confirms this trend. From 2010 to 2020, global server shipments increased from 8.9 million to 12.15 million units (Alsop), and about 1.3 billion personal computers continued to be operational as of 2016 despite a marginal downturn in sales (Statista Research Department, “Installed Base of Personal Computers”). Meanwhile, smartphone subscriptions surged from 3.6 billion in 2016 to 6.4 billion in 2022 and an anticipated 7.7 billion by 2028 (Statista Research Department, “Number of Smartphone Subscriptions”). The Web 2.0 trend of end users as both consumers and producers of digital content has consequentially fueled an exponential increase in global data volumes. Total data volume is projected to increase from approximately 41 (Taylor) to 59 ZB (IDC) in 2019 to 175 (Reinsel, Grantz, and Rydning) or 181 ZB (Taylor) by 2025.

However, a closer examination of the Jevons paradox in the Web 2.0 transition reveals a more intricate picture than what meets the eye. ICT efficiency gains and market demands manifest in surprising ways. The disparity between the proposed Kryder’s Law and actual outcomes illustrates this point. In 2005, Mark Kryder projected that digital storage capacity increases would outpace Moore’s Law (Walter) and anticipated the availability of 40-TB disks at US$40 by 2020 (Kryder and Kim, 3406). Contrary to these forecasts, the ICT sector focused less on storage volume and more on performance enhancement and power conservation, driven by the proliferation of mobile devices, cloud computing, and social media. This mismatch between predicted and actual trajectories underlines the unpredictability and complex nature of technological advancements. Theoretical models provide a framework, but real-world tech development and its impact on resources and consumption tend to diverge from expected outcomes.

From 2008 to 2023, high-end consumer and enterprise-grade SSDs experienced more than a hundredfold performance growth in both sequential and 4K random input/output operations per second (IOPS). In 2008, a standard Seagate HDD required 11,772 seconds (3.27 hours) for sequential 1-TB reads and 1,048,576 seconds (more than twelve days) for the same amount at 4K random blocks. The 2023 SSD completes these tasks in under two minutes and approximately two hours, respectively. (See the relative time measures in Figure 7.2, calculated based on the performance measures shown in Table 7.2.)7

In addition, the shift from mechanical disks to flash memory has brought considerable power efficiency gains. In 2009, Intel’s premium MLC and SLC SSDs already showcased a 27- to 213-fold decrease in energy per terabyte compared to HDDs. In 2023, Crucial’s flagship model achieved a further 6- to 10-fold reduction in electricity consumption per terabyte for sequential reads relative to Intel’s 2009 models. (See the relative energy consumption measures in Figure 7.3, calculated on the basis of energy efficiency measures shown in Table 7.3.8)

The environmental and energy costs of global ICT infrastructure similarly present a multifaceted picture. Phrases such as “more than 2 percent of global energy use” (Mullaney, 5) and “70 billion kilowatt-hours of electricity in 2016 in the United States alone” (Ensmenger, 34) demand a more nuanced interpretation. Crawford’s assertion that “the tech sector will contribute 14 percent of global greenhouse emissions by 2040” (42) carries an important caveat. This is a worst-case scenario based on an exponential fit and contingent on current trends continuing “if unchecked” (Belkhir and Elmeligi, 448), and with “large uncertainty about the lifecycle annual footprint computers . . . and displays” (458). Relative data is partially measures of ICT’s outpaced growth vis-à-vis other industries, while projections are scenario-based ranges with minimum and maximum values. The analysis also varies depending on whether the focus is solely on data center operations or also encompasses the manufacturing, installation, and usage of optical fibers, wireless networks, personal computers, mobile devices, televisions, and gaming consoles. In 2005, data centers accounted for about 1 percent of the world’s electricity, half of which was dedicated to cooling and power distribution (Koomey, 1). By 2020, data centers were responsible for 1.4 to 1.6 percent of the global carbon footprint, with communication networks and personal digital devices adding another 1.7–2 percent, while communication networks, desktops, notebooks, displays, tablets, and smartphones emitted an additional 1.7 to 2 percent.9

Eric Masanet and his team present a compelling counternarrative to the dire forecasts about data center energy consumption. Their research indicates that, as of 2020, global electricity usage by data centers remained steady, at about 1.1 to 1.5 percent, due to a concomitant rise in overall power generation to meet the lifestyle demands of the growing middle-class population, particularly in emerging economies (Masanet et al., 984; Statista Research Department, “Electricity Generation Worldwide”; Tweed). Masanet et al. remind us to be careful with “oft-cited yet simplistic analyses” (984) that overlook simultaneous trends in energy efficiency. From 2010 and 2018, there was a remarkable twenty-five-fold increase in data center storage capacity (Masanet et al., 985), but this expansion was counterbalanced by a significant ninefold reduction in the amount of energy required per unit of storage (984). While the ICT sector is a major energy consumer and contributor to emissions, it will not consume one-fifth of the world’s electricity by 2025 (Vidal), nor will Japan’s data centers gobble up the entire nation’s electricity supply by 2030 (Bawden). A balanced perspective is vital for crafting effective strategies, policies, and collective responses.

What if the objective is to drastically reduce greenhouse gas emissions in absolute terms and increase the use of renewable energy to 100 percent? The conundrum is that the ICT sector, as one of the largest carbon polluters, has responded with commitments to the transition to renewables. Take US-based Big Tech for example. Google uses a credit-matching strategy in which it purchases the same amount of renewable energy as its facilities consume. Despite Crawford’s criticism of this practice (43), it is still a positive development and a step in the right direction. Meta’s global operations have been powered entirely by renewable energy since 2018, and the company aims to achieve net zero emissions across its entire supply chain by 2030 (Parekh). In 2021, Amazon Web Services reported that 85 percent of the electricity used in Amazon’s businesses was derived from renewable sources, with the goal of reaching 100 percent by 2025 (Amazon). Recently, the cloud giant added 2.7 GW of low-carbon energy capacity in South America, India, and Poland, as well as investing in 379 renewable projects worldwide (Robinson). Despite these efforts, however, Amazon’s overall carbon footprint increased by 19 percent due to the company’s growth outperforming its green initiatives (Richardson).

The world outside the United States merits more than a passing mention. In northern Europe, Meta operates data centers in Odense and Luleå that use 100 percent renewables according to the abovementioned global commitment (Edelman; Meta), and Google powers its Hamina location with electricity from a wind farm in northern Sweden (Alley). However, one should be cautious about generalizing Nordic facilities as uniformly green. As Julia Velkova notes (672), renewable sources accounted for less than 10 percent of the electricity used at the Yandex data center in Mäntsälä, Finland, in 2018 and 2019. By contrast, Latin America demonstrates a latecomer’s advantage in building new ICT infrastructure powered primarily by renewables. Between 2010 and 2022, the renewable capacity in Latin America and the Caribbean has nearly doubled from 168 to 314 GW (Fernández, “Renewable Energy Capacity in Latin America”), with Brazil leading the charge (Fernández, “Leading Countries”). According to Boston Consulting Group, the region’s cloud market is worth US$10 billion as of 2022 and will grow at a 30 percent annual rate (Boston Consulting Group). To meet this demand, Microsoft’s Chile, Ascenty, Scala Data Centers have already built or are in the process of building carbon-neutral data centers (News Center Microsoft Latinoamérica; DF SUD; Bnamericas).

China also deserves special attention. The People’s Republic is home to one-quarter of the world’s data centers and the world’s largest emitter of greenhouse gases (BBC News). Despite its continued reliance on coal, China has reduced its coal use for electricity generation from 72.4 percent in 2005 to 56.8 percent in 2020 (Cheng) and has emerged as the leading producer of renewable energy in the world. As of 2024, China’s total renewable capacity was 1,828 GW, more than four times that of the United States (Fernández, “Leading Countries”). The issue is that renewable energy appears to be slow to reach the ICT demand concentrated on the east coast. In 2021, BloombergNEF ranked the e-commerce giant Alibaba as the largest buyer of renewable energy among Chinese companies (Alibaba Group), and Alibaba pledged to achieve carbon neutrality in all its operations by 2030 (Greenpeace East Asia). During the same year, however, Alibaba Cloud reported that only 21.6 percent of its electricity usage came from clean energy sources, which was the national average in China in 2018 (Greenpeace). Aware of this issue, the Ministry of Industry and Information Technology and National Development and Reform Commission have mandated that all new hyperscale data centers attain a power usage effectiveness (PUE) ratio of 1.25 by 2025 (Xue). China’s challenge is that that remote solar and wind generators in Gansu, Ningxia, Qinghai, Xinjiang, and Inner Mongolia must be connected to digital infrastructure in populous regions via ultra-high-voltage electricity transmission lines. This centrally planned, supply-driven strategy is met with resistance from various interest groups and will take time to gain local acceptance (Chen).

The foregoing discussion is written from the perspective of an intellectual and an environmental historian striving for a balance of ideational and materialist viewpoints. Our interest in this topic originates from many hours of musing in graduate school over a narrative of world history from the perspective of information and energy regimes. As area specialists of medieval Korea and early modern China, we approach global and comparative scholarship with an emphasis on the necessity of engaging with sources from multiple continents and in various languages. Our background in premodern history informs our tendency to stress the importance of problematizing historical temporalities in the Annales sense of short-term events, medium-term conjunctures, and enduring longue durée trends.

This chapter’s genesis is also marked by a confluence of unforeseen events. In 2021, when the editors of this Debates in the Digital Humanities volume announced the call for papers, Bitcoin’s value reached its all-time-high and the web3 community was in the midst of buoyant optimism about the imminent coming of a decentralized web. By the time we completed this chapter, however, the ICT discourse radically shifted, with an overwhelming attention given to large-scale AI and centralized infrastructure, spearheaded by the advent of OpenAI’s ChatGPT. Concurrently, Javier Cha, one of the authors of this chapter, conducted interviews and field research exploring data centers and digital archive solutions, notably at Google’s Hamina location, Microsoft Research’s Project Silica, and the Arctic World Archive in Svalbard.

Cha’s journey from Hong Kong to Svalbard required more than twenty hours of air travel. During the descent at Longyearbyen, the sobering spectacle of barren landscapes, receding glaciers, and snowless peaks revealed the stark realities of climate change in a region warming at quadruple the global average. The surreal depositing of a symbolic piqlFilm in a decommissioned coal mine within the Arctic Circle fostered a mix of humility and cautious optimism.10 The striking sight of air conditioners installed inside the Global Seed Vault, a response to the permafrost failing to maintain the required coolness during summer, raised questions about the longevity of the digital data stored in the Arctic World Archive. This poignant experience profoundly echoed what Bethany Nowviskie means by “graceful degradation, preservation, memorialization, apocalypse, ephemerality, and minimal computing” (i12) with respect to DH practice in the Anthropocene.

The path forward for DH requires an in-depth understanding of the energy dynamics in digital information processing. Advanced cold-storage solutions such as Project Silica, M-Disc, and piqlFilm prompt new inquiries about the ultra-long-term preservation of digital data, spanning thousands to millions of years. Yet, in the immediate term, the transformation of the humanities into a substantial energy consumer is a pressing concern. Minimal computing principles, while valuable, are inadequate for addressing this surge in energy demand. Unexpectedly, the recent shift to AI-optimized servers in data centers has led to a marked increase in electricity and water usage, far exceeding that of conventional Web 2.0 setups (Langley). Renewable energy sources power many of these centers, and district heating systems in Nordic and Baltic regions show how excess heat can be effectively reused (Velkova). Our own experiences with LLM training on high-end NVIDIA guzzlers, however, lead us to question the necessity of such tasks and think about future implications of these trends.

Our hope is that the DH community uses its expertise in digital technologies to bridge the gap between the research of activist groups and corporate green claims and come up with some sound solutions and responses for mitigating big data’s environmental impact. Addressing the environmental impact of digital technology will require working together with scientists, engineers, educators, and policymakers, especially in the communities most affected by climate change. The ability of DH scholars to add our deep understanding of digital media and computational methods to the traditional humanities’ strengths in critical thinking, cultural understanding, and linguistic skills positions us in a unique way to contribute meaningfully to this dialogue.

Notes

Javier Cha would like to acknowledge the support of the Innovative and Pioneering Young Researchers Scheme at Seoul National University, as well as the Seed for Basic Research for New Staff and the Dean’s Development Fund from the Faculty of Arts at the University of Hong Kong.

1. 1. This account is based on personal communication between Chinho Pak and Javier Cha, one of the authors of this chapter, on November 14, 2022 as well as an email sent by Pak to Cha on December 7, 2022.

2. 2. 3Vs (volume, velocity, and variety) is a widely accepted framework for defining the key characteristics of big data. The concept is credited to Doug Laney, who originally formulated it as 3Ds: data volume, data velocity, and data variety. For a discussion on the implications of big data for DH, see Cha.

3. 3. Digital Humanities Quarterly published fourteen articles on minimal computing in its summer 2022 issue. The guest editors, Roopika Risam and Alex Gil, provide an excellent overview of the lofty goals and complexities of pursuing minimal computing in the DH (Risam and Gil).

4. 4. Our modification has added the word “reproduction” to highlight the important differences in the energy costs of different media.

5. 5. A1: According to Smil (209, box 4.20), the approximate amount of fuel required to smelt copper is 1 kg charcoal and 1 kg slag, and the specific energy content of charcoal is 28–32 MJ/kg (12, box 1.4). Ores typically contain 0.6–1 percent copper. Using the higher purity estimate, about 100 kg of charcoal, or about 3 GJ, is needed to produce a 1-kg copper object, which is a reasonable estimate of the energy needed to produce bronze. Notably, this is significantly higher than the 90–100 MJ/kg required to produce copper using modern techniques. The high energy cost of materials completely outweighs the energy cost of human labor, which ranges from 6 to 10 MJ/person/day, depending on the size of the person and the difficulty of the labor (Smil, 19, box 1.10). A2: This figure is adapted from Smil, 19, box 1.10, which estimates the energy cost of human labor as about 7.5 MJ/person/day and accounts for the relatively light labor of writing. Assuming a median productivity of 16 folio pages/day at 7.5 MJ/person/day, the labor cost per page is approximately 0.47 MJ. A lower productivity, perhaps 2–4 pages per day, would result in a much lower efficiency at 0.94–1.88 MJ per page. In contrast, the higher scribal efficiencies observed in the late Middle Ages would have yielded a much higher efficiency of perhaps 0.2–0.3 MJ/page. Assuming a higher cost of paper or vellum than in early modern times, the paper’s embodied energy was likely between 0.3 and 0.6 MJ/page. This results in a wide range of potential energy costs for manuscripts, ranging from as low as 0.5 MJ/page to as high as 2.5 MJ/page, with human labor accounting for 60–75 percent. A3: This figure follows Kai-wing Chow’s estimates of 200–400 characters per folio side and 100 characters/person/day for carving woodblock (35–38). Smil (16, box 1.8) calculates embodied energy to be between 1 and 3 MJ/kg for lumber and 23 and 35 MJ/kg for paper. Accounting for the lighter labor and adapting from Smil, 19, box 1.10, the energy cost of human labor turns out to be approximately 7.5 MJ/person/day. Using these numbers, the energy cost of one folio page is 15–30 MJ/page for labor (2–4 days of work at 7.5 MJ/day), 0.75 MJ/page for woodblock (0.5 kg/page at 1.5 MJ/kg), and 0.15 MJ/page for paper (5 g/page at 30 MJ/kg). B3: Assuming that a worker could produce 1,500 pages/day (Chow, 70), the energy expended works out to 0.005 MJ/page. Lucien Febvre and Henri-Jean Martin (132) provide a higher estimate of up to 3,350 impressions/day for printing with movable type, resulting in an energy cost of slightly more than 0.002MJ/page. With several orders of magnitude greater production rates, industrial printing would have truly marginal energy costs for human and machine output. In every instance, the materials cost dominates at 0.15 MJ/page for paper and a somewhat lower cost for ink (perhaps 0.05 MJ/page as a high estimate). The total is about 0.200 MJ/page or 200 KJ/page. A4: Compositors could complete 1 to 3 sets per day, which equates to approximately 1 to 6 pages (i.e. 1 set of 4 pages or 2–3 sets of 2 pages) (Febvre and Martin, 131–32). This results in labor costs of 1.25 to 1.91 MJ/page for setting forms (excluding checking) and 0.002 to 0.004 MJ per page for printing. Paper and ink cost 0.2 per page in terms of materials. Because type could be reused, it is much more difficult to estimate the cost of fixed materials, but they were probably marginal compared to the cost of the labor and paper. The total comes to somewhere between 1.45 and 2.10 MJ/page. A5: This figure is from Burr. A6: Assuming a computer operating at 75 w, five minutes of typing would require 22.5 kilojoules of energy. This represents a high-end estimate of power consumption. In actuality, while Ian Miller was typing this page, his 60-w laptop consumed less than 5 w on average, or 1.5 kilojoules of energy consumed in five minutes. If this is performed on cloud-hosted documents, the cost could be multiplied by two to four, or 3 to 6 kJ for five minutes. A human typist would consume an astoundingly similar amount of energy as a computer running full tilt. 7.5 MJ/day translates to 5.2 kJ/min, or approximately 26 kJ for five minutes. A slower typist taking ten minutes would increase these figures by twofold. On Miller’s laptop, it takes between 29 and 64 kJ to type a page of text, depending on my typing speed and power consumption. An extremely high-energy use case could reach 100 kJ/page. B6: For instance, it took approximately two seconds to copy and load a one-page document on Google Drive, which caused the laptop’s power consumption to spike to approximately 20 w. Assuming that this resulted in a similar power surge on the server side, it took approximately 80 J to generate the document. Longer documents do not increase the amount of time required in a linear fashion, so their per-page energy consumption decreases to probably just a few joules.

6. 6. The minimal charge requirement of SSDs for long-term data storage is poorly understood. In 2015, ExtremeTech reported that SSDs removed from a power source could begin to lose data after only 20 to 105 weeks (Hruska). This claim turned out to be a misinterpretation of the data presented by Alvin Cox of JEDEC, a semiconductor trade organization and standardization body (Ung). In addition, SSDs have finite read/write cycles, which can shorten its lifespan, but this characteristic is unrelated to energy use.

7. 7. The performance measures for each HDD or SSD in this table are from PassMark Software’s Hard Drive Benchmarks—Seagate ST31000333AS 1TB HDD, https://www.harddrivebenchmark.net/hdd.php?id=112; Intel X25-M 80GB MLC SSD, https://www.harddrivebenchmark.net/hdd.php?id=11053; Intel X25-E 32GB SLC SSD, https://www.harddrivebenchmark.net/hdd.php?id=1651; Crucial T700 2TB SSD, https://www.harddrivebenchmark.net/hdd.php?id=34572. Among various metrics for measuring the performance of HDDs and SSDs, the distinction between sequential reads and 4K random IOPS is crucial. Sequential reads measure the speed at which data is read from the storage device when the data is arranged contiguously or in order, such as when transferring a large movie file from one location to another. On the other hand, 4K random IOPS is a performance measure for tasks dealing with small (i.e., 4K) data chunks located randomly across HDDs and SSDs, which is useful for understanding scenarios where the workload involves a lot of small, scattered day-to-day tasks, such as web browsing, multitasking environments, and server applications where access patterns involve numerous small files.

8. 8. The estimated energy consumed per terabyte is calculated by multiplying the average wattage by the time in seconds required to read 1 TB. For example, Crucial T700 2TB SSD, with an average power consumption of 6.7 w, requires 107.75 seconds to read 1 TB sequentially, returning 722 J per TB. The individual power requirement figures have been obtained as follows: Seagate ST31000333AS 1TB, average power requirement 10.2 w (source: Schmid and Roos, “Tom’s Winter 2008,” 15); Intel X25-M 80GB MLC SSD, average power requirement 1.69 w (source: Schmid and Roos, “Intel X25-E,” 7; Intel X25-E 32GB SLC SSD, average power requirement 1.1 w); Crucial T700 2TB SSD, average power requirement 6.7 w (source: “Crucial T700 2 TB”).

9. 9. This figure is obtained by multiplying 3.1 to 3.6 percent of total relative carbon footprint by 45 percent and 55 percent, respectively (see Belkhir and Elmeligi, 457).

10. 10. piqlFilm is a long-term digital preservation medium developed by the Norway-based company Piql. It records textual and visual information by printing directly onto film and binary data as high-definition QR codes on film coated with silver halides, engineered to resist degradation. Piql also operates the Arctic World Archive in Svalbard, where reels of piqlFilm are sealed in containers and stored inside a decommissioned coal mine with the intention of being maintained at optimal conditions by the surrounding permafrost. GitHub is archived in the Arctic World Archive.

Bibliography