PaperlayerView exact revision on arXiv

Source-linked AI summary

Embodied Emissions of 2.5D and 3D Scaling in Chipletized GPUs for Low-Carbon AI

Hannah Watson, Arjun Singh, Jennifer Stephenson, Job Soethoudt, Jonas Svedas, Debjyoti Bhattacharjee, James Myers

arXiv:2610.32830v1cs.AR

TL;DR

Growing AI workloads increase the emissions stakes of scaling GPUs, so this paper models embodied emissions for chiplet-based GPUs scaled in 2.5D and 3D. The 3D scaled system has an 89% embodied-emissions increase, compared with 141% for 2.5D, while performance gains are modest.

  • Problem

    GPU scaling for growing AI workloads raises the climate impact of hyperscale data centers, motivating comparison of embodied emissions across packaging choices.

    Source: §I · p. 1

  • Method

    The study uses imec.netzero to estimate embodied emissions for a chiplet-based GPU scaled in 2.5D and 3D with increased compute and HBM capacity.

    Source: §II · p. 1, §I · p. 1

  • Results

    89% higher embodied emissions for the 3D scaled system compares with 141% for the 2.5D scaled system, alongside only 6.8–13.5% performance gain.

    Source: §V · p. 2

  • Takeaways & Limitations

    3D integration is an attractive option to limit embodied-emissions increases by building compact systems without significantly impacting performance.

    Source: §V · p. 2

    Reduced yields from increased HBM stack height and thermal-management requirements may become bottlenecks that limit 3D technology adoption.

    Source: §V · p. 2

Abstract

from arXiv · show

GPU compute resources are scaling to support growing AI workloads, leading to further rises in the climate change impact of hyperscale data centers that rely on high-performance GPUs. To model how packaging decisions impact the carbon footprint we have estimated the embodied emissions (EE) of a chiplet-based GPU that has been scaled in 2.5D and 3D using the imec.netzero modelling platform. We doubled the number of compute units and high bandwidth memory (HBM) capacity in the 2.5D and 3D scaled systems and found that 3D scaling results in an 89% increase in EE, as compared to an increase of 141% for 2.5D scaling, positioning 3D scaling as an attractive solution to limit the embodied emissions due to GPU scaling.

I. INTRODUCTION

Growing AI workloads are increasing demand for hyperscale data-center capacity, while GPUs contribute substantially to emissions. The paper investigates how 2.5D and 3D scaling choices affect embodied emissions.

I. INTRODUCTION

  • Data-center capacity demand is predicted to nearly triple by 2030, with 70% associated with AI workloads.Source: §I · p. 1
  • GPUs accelerate AI workloads and contribute substantially to emissions because of their growing design complexity and power requirements.Source: §I · p. 1
  • The study compares 2.5D and 3D scaling of a chiplet-based GPU by increasing compute units and HBM per GPU.The design is inspired by the AMD MI300X GPU.Source: §I · p. 1

II. EXPERIMENTAL SETUP

The study estimates embodied emissions with a bottom-up manufacturing model and calculates die and assembly yields using specified assumptions. Its modeled packaging flows exclude emissions from extra DRAM-die thinning in HBM stacks.

II. EXPERIMENTAL SETUP

  • imec.netzero estimates embodied emissions using a bottom-up model of a virtual high-volume IC manufacturing fab across semiconductor components and technology nodes.Components are specified with technology parameters, including process node and wafer-level packaging flows.Source: §II · p. 1
  • Current modeled wafer-level packaging flows omit emissions from extra thinning of DRAM dies in HBM stacks.Source: §II · p. 1
  • Logic and interposer die yields use the Murphy model with die area, defect density of 0.1 defects/cm2, and critical-area fractions of 100% and 15%, respectively.The 100% fraction applies to logic dies and 15% to the passive interposer.Source: §II · p. 1
  • The model assumes uniform defect sizes, fixed yields for DRAM and dummy silicon dies, and excludes reticle-stitching and warpage effects.The authors note passive interposer yield may be higher because only larger defects cause failures for large feature sizes.Source: §II · p. 1
  • Assembly yield assumes known-good dies and subpackages are assembled, testing occurs after assembly, and packaging yield is 99% per die and subpackage attachment.Assembly connectivity failures can result from TSV fabrication errors or misalignment, and can cause loss of all subcomponents and subpackages.Source: §II · p. 1

A. Baseline System

The baseline is a chiplet GPU combining 2.5D and 3D integration, with compute and HBM components arranged in subpackages. HBM DRAM dies are the largest embodied-emissions contributor.

A. Baseline System

  • HBM DRAM dies account for 53.2% of baseline embodied emissions, followed by logic dies at 28%.Source: A. Baseline System · p. 1
  • The baseline combines one passive silicon interposer, four I/O dies, eight accelerator compute dies, and eight 12-Hi HBM3 stacks.The HBM stacks are microbumped onto HBM3 logic dies; accelerator dies are hybrid-bonded to I/O dies.Source: A. Baseline System · p. 1
  • The system mixes 2.5D and 3D integration, with IOD-XCD and HBM 3D stacks serving as its subpackages.Dummy silicon components provide structural stability and thermal dissipation.Source: A. Baseline System · p. 1

B. Scaling Compute Resources In 2.5D

The 2.5D design doubles HBM capacity and compute dies in a package twice the baseline size. Its larger interposer and doubled subpackage count reduce yields, while embodied emissions rise substantially.

B. Scaling Compute Resources In 2.5D

C. Scaling Compute Resources In 3D

The 3D-scaled system adds compute and HBM capacity with a smaller package increase than the 2.5D system, and its modeled embodied emissions are lower. The comparison attributes the difference largely to fewer IODs, a smaller interposer, and fewer subpackages.

C. Scaling Compute Resources In 3D

  • The 3D system adds 8 XCDs and uses 12 stacks of 16-Hi HBM, increasing package size by 25% to match 2.5D compute and HBM capacity.Its 16-Hi stacks align with the upcoming HBM4 specification.Source: C. Scaling Compute Resources In 3D · p. 2
  • The 3D system’s IOD-XCD subpackage yield falls to 95.1%, HBM subpackage yield to 84.3%, and GPU assembly yield to 85.1%.The estimate excludes yield losses from the increased risk of warpage or cracking associated with additional die thinning for 16-Hi stacks.Source: C. Scaling Compute Resources In 3D · p. 2
  • 3D scaling raises EE 88.5% over baseline, while 2.5D EE is 28% higher than 3D.The 3D system’s 88.5% increase is reported in the detailed scaling passage.Source: D. 2.5D vs. 3D Scaling Comparison · p. 2, C. Scaling Compute Resources In 3D · p. 2
  • The 2.5D system’s EE is higher largely because it requires more IODs, a larger interposer, and more subpackages; its interposer is 60% larger than 3D’s.The 2.5D system has double the number of IODs, while 3D stacking allows fewer IOD-XCD subpackages.Source: D. 2.5D vs. 3D Scaling Comparison · p. 2

IV. PERFORMANCE

The systems have equal projected FLOPs and HBM capacity, but differ in bandwidth and network topology. In the evaluated training workloads, the 2.5D system achieves only modest speedup despite its bandwidth advantage.

IV. PERFORMANCE

  • Despite high compute–communication overlap in the evaluated training workloads, 2.5D achieves only 6.8–13.5% speedup.The workloads are evaluated using Hespas after partitioning each system around its IODs.Source: §IV · p. 2
  • The systems are iso-FLOPs and iso-HBM capacity, but 2.5D provides higher HBM bandwidth because it has more HBM packages.Its eight IODs form two switched, fully connected four-IOD groups, whereas 3D has a fully connected topology among four IODs.Source: §IV · p. 2
  • The 3D system is more effective for workloads with lower compute–communication overlap.The passage contrasts its fully connected four-IOD topology with the two switched groups in 2.5D.Source: §IV · p. 2

V. CONCLUSIONS

The conclusion compares embodied emissions from 2.5D and 3D GPU scaling, finding a smaller increase for 3D alongside modest performance gains. It identifies yield and thermal-management requirements as potential constraints on 3D adoption.

V. CONCLUSIONS

  • 2.5D scaling increases EE by 141%, compared with 89% for 3D, while performance gains are 6.8–13.5%.The conclusion presents 3D integration as an attractive option for limiting EE increases without significantly affecting performance.Source: §V · p. 2
  • Reduced yields from increased HBM stack height and thermal-management requirements may limit adoption of 3D technology.The conclusion also notes that longer wire lengths in 2.5D may degrade performance.Source: §V · p. 2
  • Assessing EE during chip design can help identify carbon-efficient solutions for low-carbon AI.Source: §V · p. 2

Baseline 2.5D Scaled

The supplied figures and table describe comparisons across baseline, 2.5D-scaled, and 3D-scaled systems, covering EE, assembly yield, compute resources, network configuration, and training-workload performance.

Baseline 2.5D Scaled

  • Normalized EE is 141% higher for 2.5D and 89% higher for 3D than baseline.The figure also compares EE by category across baseline and both scaled systems.Source: Figure 5 · p. 3
  • Figure 6 compares EE for baseline and both scaled systems across the highest-contributing subcomponents.Source: Figure 6 · p. 3
  • Table II summarizes compute resources and normalized EE; projected peak FLOP/s scales published MI300X values by compute-unit count.Source: Table II · p. 3
  • GPU assembly yields are 88.6% for baseline, 78.6% for 2.5D scaled, and 85.1% for 3D scaled.Figure 7 compares the impact of assembly yield on EE.Source: Figure 7 · p. 4
  • Figures 9 and 10 compare the systems’ IOD network topologies and training-workload memory-bandwidth use, FLOP/s use, and runtime.Figure 9 reports higher memory pressure for 3D, while 2.5D has slightly improved compute utilization and lower runtime.Source: Figure 9 · p. 4, Figure 10 · p. 4
  • The supplied table captions describe compute and memory partitioning around IODs, network configurations, and training models exported in StableHLO format.Source: Baseline 2.5D Scaled · p. 4
Loading 2610.32830v1…