SCARIF: Towards Carbon Modeling of Cloud Servers with Accelerators

Overview

The authors collected 96 vendor product-carbon-footprint reports for servers from Dell, HP and Lenovo, all produced with the PAIA LCA tool, and built a linear model of embodied carbon. It predicts a server's embodied carbon from total CPU core count, SSD, HDD and memory capacity, and release year. Bottom-up tools such as ACT leave out mainboard PCBs, connectors, PSUs, chassis and other peripherals, so they underestimate whole-server embodied carbon several-fold.

In vendor reports covering a 4- or 5-year lifetime, the operational share of server life-cycle carbon is 70.2-90.2 percent (Dell), 66.2-93.7 percent (HP) and 39.4-97.0 percent (Lenovo); the manufacturing share is 9.5-27.7, 5.4-28.4 and 2.9-59.5 percent. Transport and end-of-life are usually below 1 percent. The SCARIF model is E = K1*#CPU cores + K2*SSD_GB + K3*HDD_GB + K4*Mem_GB + K5*(year-2000) + D, with K1 = 5.01 kgCO2e per core, K2 = 0.16, K3 = 0.04, K4 = 0.95 kgCO2e/GB, K5 = 83.08 kgCO2e/year and D = -1100 on HP data.

In validation, predictions fall within 0.4 sigma of Dell's reported values for 37 servers, with an average error within 0.15 sigma; the average error on 40 Lenovo reports is 50 percent. For a Dell R740, ACT and the vendor PAIA LCA differ 2.26x in total; components ACT omits add 136.3 kgCO2e, and the mainboard PWB alone is 108.68 kgCO2e. An R740 with 2x Xeon 8180, 1 TB HDD and 64 GB memory has 1993.72 kgCO2e embodied, or 2158.53 kgCO2e with one V100 GPU.

In the case study, the carbon breakeven for upgrading R740+V100 to R750+A100 is 4.2 years (TX), 4.6 (AZ), 7.8 (CA) and 9.8 (NY), using grid intensities of 0.438, 0.395, 0.234 and 0.188 kgCO2e/kWh; this exceeds the roughly 4-year vendor lifetime. ACT's lower embodied estimate would give a breakeven under 1 year and reverse the decision. The document is a preprint, not peer reviewed, and the model is an approximate estimate from vendor data; the error is large on Lenovo data.

Key points

1. In vendor reports, the operational share is 70.2-90.2 percent (Dell), 66.2-93.7 percent (HP) and 39.4-97.0 percent (Lenovo); transport and end-of-life are usually below 1 percent.
2. Total embodied carbon of the highest/lowest servers: Dell 1782/1133, HP 3880/423, Lenovo 15593/585 kgCO2e.
3. SCARIF: K1 = 5.01 kgCO2e per core, K2 = 0.16 (SSD), K3 = 0.04 (HDD), K4 = 0.95 (memory) kgCO2e/GB, K5 = 83.08 kgCO2e/year, D = -1100 on HP data.
4. Validation: within 0.4 sigma of Dell's reported values for 37 servers, average error within 0.15 sigma; average error on 40 Lenovo reports is 50 percent.
5. For a Dell R740, ACT and the vendor PAIA LCA differ 2.26x; omitted components add 136.3 kgCO2e, the mainboard PWB alone is 108.68 kgCO2e.
6. Breakeven for upgrading R740+V100 to R750+A100 is 4.2 (TX), 4.6 (AZ), 7.8 (CA) and 9.8 (NY) years, exceeding the roughly 4-year vendor lifetime.

Using it in design

• Add an embodied-carbon estimator per server SKU using the SCARIF linear model (5.01 kg per CPU core, 0.16 kg/GB SSD, 0.04 kg/GB HDD, 0.95 kg/GB DRAM, 83.08 kg x (year-2000), intercept -1100), and tag the result as an approximate, vendor-calibrated value.
• Let vendor PCF report values override the estimate.
• Include a refresh-breakeven calculator that weighs a new server's embodied carbon against its annual energy savings at the site's grid intensity, and warn when breakeven exceeds the planned server lifetime.
• Warn users that chip-only (ACT-style) estimates can understate whole-server embodied carbon by more than 2x.

Who should read it

Sustainability and capacity planning engineers weighing embodied carbon in server refresh decisions and Scope 3 accounting.

Publication details

Authors

Shixin Ji; Zhuoping Yang; Xingzhen Chen; Stephen Cahoon; Jingtong Hu; Yiyu Shi; Alex K. Jones; Peipei Zhou

Publisher

arXiv preprint (cs.DC)

Year

2024

Document number

arXiv:2401.06270

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2401.06270

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation