Technical Assessment

Decisions

Final Decision: ACCEPTED

1. Are the codes to be used suitable for Leonardo's architecture? Yes. The project utilizes a standard deep learning stack (PyTorch, MONAI, NumPy, CUDA) that is fully compatible with the NVIDIA A100 GPUs on the Leonardo BOOSTER partition. The use of SIREN and Persistent Homology libraries is also compatible with the system's GPU and CPU architecture.

2. Is the code installed on Leonardo? Partially. While the specific research code (CARAVEL) is not a pre-installed system module, the applicant has explicitly verified that a functional PyTorch DDP implementation of the framework has already been tested and executed on Leonardo. The underlying libraries (PyTorch, CUDA) are available via the cineca-ai module.

3. Are the proposed targets feasible within the allocation time? Yes. The requested 4,500 node hours are within the fixed cap for Development calls. The project's objectives—optimizing MoE routing, implementing topology-aware supervision, and benchmarking scaling efficiency—are realistic for the 12-month duration and the requested resources.

4. Technical data quality and conclusions The technical data provided is of high quality. The applicant demonstrates a deep understanding of both the scientific challenges (vascular topology) and the HPC bottlenecks (MoE routing overhead, DDP synchronization, and GPU memory). The workplan is clear, and the performance targets (e.g., $\ge 60\%$ strong-scaling efficiency) are measurable and realistic.

5. Reasons for final classification The application is ACCEPTED because it meets all technical requirements of the Development call. The resource request is exactly at the cap (4,500 node hours) and does not exceed it. The software stack is fully compatible, the technical approach is sound, and the project is already in a functional state on the target hardware.

Summary Table

Application Metadata

• Application ID: EHPC-DEV-2026D11-264
• Evaluation Date: 2026-10-19
• Requested Resources
• partition: Leonardo BOOSTER
• node hours: 4,500
• project duration in months: 12
• storage: 5,000 GB
• max number of nodes: Not explicitly stated (derivable as $\le 256$ based on parallelism)
• Suggested Granted Allocation (if different from requested): as requested

Project and Executive Summary

The project aims to develop and optimize a computational framework for learning population-level cerebrovascular atlases from 3D Magnetic Resonance Angiography (MRA). It builds upon the VITAL framework, utilizing continuous neural implicit representations (SIREN) and a Mixture-of-Experts (MoE) architecture. The primary technical goals are to incorporate topology-aware supervision via persistent homology to ensure vascular connectivity and to optimize the distributed training workflow on HPC systems.

The workplan is technically sound, focusing on resolving specific bottlenecks such as MoE routing overhead and GPU memory consumption. The project is highly feasible, as the team has already established a functional multi-node pipeline on Leonardo. All resource requests are within the limits of the Development call, and the software stack is fully compatible with the BOOSTER partition.

Sub-Agent Evaluations

1. Code Compatibility Assessment

Verdict: ACCEPTED Summary: The software stack is standard and fully compatible with the A100 architecture. Evidence: PyTorch, MONAI, and CUDA are natively supported on Leonardo BOOSTER.

2. Feasibility Assessment

Verdict: ACCEPTED Summary: Resource requests are within all call-specific caps. Evidence: - Node hours: 4,500 / 4,500 (100% of cap) - Storage: 5 TB / 10 TB (50% of cap) - Data Transfer: 2 TB (within daily reference) - Memory: 256 GB total usage is well within the 512 GiB per node limit.

3. Technical Details Assessment

Verdict: ACCEPTED Summary: High-quality technical proposal with a clear optimization path. Evidence: The project identifies specific HPC bottlenecks (MoE routing, DDP synchronization) and proposes concrete solutions (batched execution, precomputation).

Synthesis & Final Reasoning

The evaluation across all three technical dimensions (Code Compatibility, Feasibility, and Technical Details) resulted in an ACCEPTED status.

1. Code Compatibility: The stack is standard for AI on Leonardo. The applicant's previous successful tests on the system remove any ambiguity regarding installation or runtime compatibility.
2. Feasibility: The request is perfectly aligned with the Development call's constraints. The node-hour request is exactly at the 4,500 cap, and storage/data transfer requests are conservative.
3. Technical Quality: The proposal is scientifically and technically robust, moving beyond a simple "request for resources" to a detailed optimization plan with measurable KPIs.

Following the fail-safe logic: All sub-agents returned ACCEPTED $\rightarrow$ Final Decision: ACCEPTED.

Recommendations (if applicable)

• The applicant is encouraged to use Apptainer (Singularity) containers to ensure the reproducibility of the MONAI/PyTorch environment.
• Use of NVIDIA Nsight Systems is recommended to further profile the MoE routing overhead.
• Ensure the use of torch.cuda.amp to maximize the efficiency of A100 Tensor Cores.

Confidence Level

High. All technical requirements are met, and the applicant has already demonstrated functional code on the target system.