ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsTopic · imecQuick take2 min readDisconfirmAI Editorial

imec and Vrije Universiteit Brussels Model: GPU 3D Packaging Expansion Implies Embodied Emissions Increase 52 Percentage Points Lower Than 2.5D

A team from imec and Vrije Universiteit Brussels used the netzero modeling platform to estimate the embodied emissions of chiplet-based GPUs under 2.5D and 3D scaling scenarios. With compute units and High Bandwidth Memory (HBM) capacity doubled, the embodied emissions increase for 3D scaling was 89%, compared to 141% for 2.5D; the research team thus positions 3D stacking as an attractive solution for limiting the embodied emissions associated with GPU expansion. These results come from a preprint paper and represent modeling estimates rather than independent measurements.

2026-11-03 04:00 GMT+8

AI computing power demand continues to grow, and carbon emissions from hyperscale data centers are receiving increasing attention. While operational energy consumption is the most discussed aspect, the energy and materials consumed during chip manufacturing itself—known as embodied emissions—also constitute a significant environmental cost. As GPUs transition from single chips to chiplet architectures and scale computing power through 2.5D or 3D packaging, there has previously been a lack of quantitative comparison regarding the differences in embodied emissions between these packaging routes.

Researchers from imec and Vrije Universiteit Brussels filled this gap in a preprint paper released in October 2026. Using imec’s netzero modeling platform, they modeled a chiplet-based GPU scaled in both 2.5D and 3D configurations, doubling the number of compute units and HBM capacity, and then estimated the increase in embodied emissions for both approaches.

The modeling results show that the embodied emissions increase for 3D scaling is 89%, while for 2.5D scaling it is 141%. Based on this, the research team positions 3D stacking as an attractive solution for limiting the embodied emissions brought about by GPU expansion.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

It is important to emphasize that this is a preprint paper that has not yet undergone peer review, and the results stem from modeling estimates rather than actual product measurements. The specific percentages depend on the comparative conditions set by the study—namely, the simultaneous doubling of compute units and HBM capacity—and actual product configurations may yield different proportions. Furthermore, embodied emissions are only one dimension of GPU expansion decisions; factors such as performance, yield, thermal management, and cost are equally important.

For AI infrastructure planning and the semiconductor packaging industry, this study provides a quantifiable reference point: when pursuing computing power expansion, the choice of packaging architecture significantly impacts embodied emissions, making the 3D route worth considering from an environmental cost perspective.

Sources

  • semiengineering.com↗
Topic

imec

imec 联合布鲁塞尔自由大学通过 netzero 平台建模发现,在加倍计算单元和 HBM 容量时,GPU 采用 3D 堆叠封装的隐含碳排放增幅为 89%,显著低于 2.5D 方案的 141%【背景】。这一结论旨在将 3D 集成定位为限制 …

2 steps1 piecesUpdated 2026-11-03 10:02

How it got here

  1. 2025-12-08imec 在 IEEE IEDM 发布首个 3D HBM-on-GPU 热 STCO 研究,显示通过优化可将峰值温度从 141.7°C 降至 70.8°C。imec-int.com ↗
  2. 2026-11-03imec 与布鲁塞尔自由大学团队发表预印本论文,估算显示 3D 扩展的隐含碳排放增幅(89%)比 2.5D(141%)低 52 个百分点。semiengineering.com ↗

What is still unsettled

  • 3D 堆叠在实际大规模生产中的组装良率是否会抵消建模中预测的碳排放优势?
  • STCO 方案中提到的 28% 工作负载性能惩罚在不同 AI 训练任务中的具体影响范围是多少?
  • 除了频率缩放和双面冷却,是否有其他不牺牲性能的散热策略能进一步降低 3D 架构的热风险?
Read the full topic →
Continue researching →