CarbonSet is a carbon footprint (CFP) dataset contains over 1600 processors. This dataset includes data for desktop and datacenter CPUs and GPUs across multiple metrics - design metrics (chip area, technology node, transistors, TDP), performance metrics (Geekbench-OpenCL,PassMark), and sustainability metrics (ECFP, OCFP, total CFP). It also offers composite metrics like Performance per CFP and ECFP per unit area (ECFPA) for deeper tradeoff-based trend analysis.
There are 4 main aspects of CarbonSet that are briefly described in the sub-sections below:
CarbonSet.csv is the main dataset containing multiple CFP enhanced metrics stored in dataset/folder. The CFP values are modeled using our enhanced ECO-CHIP.
For simplicity, we provide dataset/CarbonSet.ipynb as a fast-track option to generate the complete CarbonSet.csv by interacting with our updated ECO-CHIP framework.
We primarily use ECO-CHIP to model the CFP of processors. We updated ECO-CHIP to make it probabilistic. This means that input parameters to ECO-CHIP can now be distributions (probability distributions or kernel density estimations) instead of fixed values. This change was inspired by the observation that several parameters that contribute to CFP have uncertainty in them. For example, parameters of semiconductor manufacturing, such as defect density, are not fixed. Instead, they vary due to various factors such as geo-locations, different implementations of a process node and later process improvements, etc. Therefore, in our enhanced ECO-CHIP, the output CFP is a distribution to better cover the wide range of possible values.
You can find the source code of the enhanced ECO-CHIP in eco_chip_enhanced folder, which contains all the required parameters and modeling functions we implemented.
In this repo, we provide 2 function modules eco_chip_enhanced and utils. eco_chip_enhanced is the main module that calculates the CFP values and generates architecture specs. utils has some reuseable functions that generate probabilistic sequences and wrapper functions to call eco_chip_enhanced. Detailed use cases are shown in dataset/CarbonSet.ipynb.
We collected performance metrics of flagship processors over the last decade for both desktop series and datacenter series. The collected datasets can be found in flagship_trends/dataset. We provide a historical view of processors design trend in sustainability-performance balance using aforementioned metrics. You can also find the scripts to generate corresponding images for better visualization in the same folder.
-
How has the AI boom impacted CFP ?
We evaluated the total CFP increase trend in datacenter GPUs, since the modern AI boom (2017 timeframe), mostly run on NVIDIA datacenter GPUs. We do a rough estimation of the units of NVIDIA GPUs sold from publically available reports. We observe that while the performance per unit CFP has increased, the total CFP has significantly increased owing to significant increase in GPU shipments.
You can find the plots, datasets and scripts in
case_studies/AI_Boom -
Is manufacturing cost ($) a valid proxy for Embodied CFP ?
Some studies claim that the manufacturing cost ($) may be sufficient to approximate the chip embodied CFP, so we don't need detailed modeling of CFP. However, in our study, we collected the manufacturing cost and selling price of flagship datacenter GPUs in 28nm, 14mn, 7nm and 5nm and compared them to the Embodied CFP numbers. The plot shows Manufacturing cost of selling price are not valid proxies of ECFP.
You can find the plots, datasets and scripts in
case_studies/ECFP_proxy -
How much must processor lifetime increase by to effectively amortize the Embodied CFP ?
We sweep the lifetime and idle time of selected flagship processors and further generate a heat map of Embodied CFP vs. Operational CFP to see when Embodied CFP exceeds Operational CFP and vice-versa. This heat map helps customers to make environment-friendly purchase decisions by considering amortizing the Embodied CFP based their own process lifetime and idle time.
You can find the plots, datasets and scripts in
case_studies/ECFP_amortization -
Are Chiplet architecture always more sustainable than monolithic architecture ? To anser above question, we analyze the manufacturing CFP of AMD's flagship chiplet CPUs over the past five years, varying chiplet counts within a fixed total chip area to determine the optimal configuration for each processor. Assuming an even distribution of total chip area across all chiplets, manufactured using the same process node, we also examine how manufacturing CFP scales across different chiplet configurations to assess whether chiplet architecture consistently provides a more sustainable solution.
You can find the plots, dataset and scripts in
case_studies/AMD_chiplet_cfp_studyandcase_studies/Area_chiplet_cfp_study.
Python >= 3.8, numpy, matplotlib, pandas, json