Invitation to Participate: NVIDIA Python Workshop at CERN

[Ianna Osborne]

Dear colleagues,

I hope you are doing well. I am reaching out to the tracker community, the analysis community, and the HSF Python ecosystem to share an opportunity that may be of interest.

I am currently organizing a workshop at CERN in collaboration with the NVIDIA Python team. The goal is to bring their experts on GPU-accelerated Python workflows to CERN for a hands-on, technically focused training tailored to our environment and use cases.

Before finalizing the structure and content with NVIDIA, I would like to gather input from the communities who would benefit most. Your feedback will help ensure the workshop reflects real needs across experiments and software domains.

Specifically, it would be helpful to know:
- Your level of interest in attending such a workshop
- Topics or tools you would like to see covered (e.g., CuPy, Numba, RAPIDS, CUDA Python, profiling, multi-GPU workflows)
- Any CERN-specific workflows or datasets you would like included
- Preferred workshop format (lecture, hands-on, hack session)
- Any constraints or considerations we should keep in mind

Once I collect your input, I will consolidate it and work with NVIDIA to shape a program that is relevant and impactful for our community.

If you are interested, please reply with your thoughts or feel free to forward this message to others who may want to participate.

Thank you in advance for your feedback. I look forward to hearing from you and building a workshop that supports our shared Python and GPU development efforts.

Best regards,
Ianna

Sent from my iPhone

[Agamya Samuel]

Dear Ianna (and the tracker, analysis, and HSF Python communities),

This is a timely and high-potential initiative. CERN’s push toward heterogeneous computing (CPUs + GPUs + emerging accelerators) aligns perfectly with the HSF’s Python-centric analysis ecosystem. GPU-accelerated Python tools can dramatically speed up columnar analysis, irregular data processing (e.g., via Awkward Array), Monte Carlo simulations, tracking/ML inference, and end-to-end workflows — areas where HEP datasets are growing exponentially.

A fruitful workshop is one that delivers immediate productivity gains, fosters lasting adoption, sparks collaborations, and scales beyond the event. This requires deliberate design across preparation, content, delivery, and follow-through. Below are a few ideas that I could think of from multiple angles: pedagogical effectiveness, technical relevance to CERN/HEP, logistical realities, inclusivity/engagement, measurement of success, and potential challenges with mitigations.

1. PRE-WORKSHOP PREPARATION: ALIGN ON NEEDS AND LOWER BARRIERS

To ensure relevance and high attendance/impact:

• Targeted needs assessment (beyond the initial email): Distribute a short, structured survey segmented by role (tracker/reco, analysis, simulation, ML) and experience level. Ask about top pain points in current Python workflows and what GPU access people already have (lxplus-gpu, HTCondor, SWAN, personal machines, cloud). There are numerous topics that can be taken up during the Workshop, but if none of them solves any real world problems faced by the HSF Community, the Workshop will be of not much Impact, and will be a one off event.
• Prerequisites and onboarding materials: Provide 1–2 weeks of self-paced prep via a dedicated GitHub or CERN GitLab repository. Include setup guides for CERN environments, Jupyter notebook templates using SWAN or JupyterHub with GPU kernels, a “HEP GPU readiness” checklist, and short videos on key concepts. (This is to make sure when the Workshop begins, everyone is in the same starting position - and everyone benefits equally from the Workshop.)
• Participant selection and cohorting: Cap at 40–60 for interactivity (in-person at CERN + hybrid option). Create mixed-ability breakout groups. Offer travel support for early-career researchers from smaller institutes to boost diversity (if possible).

Important nuance: Not everyone has easy GPU access, so plan for a “CPU-fallback” mode in all exercises so remote or CPU-only participants can still engage fully.

2. WORKSHOP STRUCTURE AND FORMAT: BLEND THEORY, PRACTICE, AND APPLICATION

A 2–3 day event (e.g., 1.5 days core + 0.5–1 day hack/extension) strikes the best balance.

Recommended hybrid format (60% hands-on):

Why this mix works: Pure lectures lead to low retention; pure hacks overwhelm beginners. Include “hackathon rules”: participants bring their own dataset/workflow (tracking algorithm, analysis ntuple, simulation loop, etc.) and get mentor help porting it.

Core tools/topics to prioritize (tailored to HEP):

• CuPy (drop-in NumPy replacement — huge win for columnar analysis)
• Numba (JIT kernels + CUDA Python for custom event processing)
• RAPIDS (GPU DataFrames/cuDF for fast ETL)
• Profiling (Nsight, CUDA Python tools)
• Multi-GPU (Dask + RAPIDS or NCCL)
• CERN/HEP-specific integrations: Awkward Array + GPU backends, uproot → Awkward → CuPy/Numba, Coffea/processor frameworks on GPU, pyhf or zfit on GPU, domain examples from tracking (Patatrack style), beam dynamics (Xsuite), Monte Carlo transport.

3. HANDS-ON PRACTICALITIES AND RESOURCE CONSIDERATIONS

• Environment: Run everything on lxplus-gpu or provisioned HTCondor GPU slots. Provide pre-built Singularity/Apptainer containers for reproducibility. NVIDIA can supply cloud credits for overflow participants.
• Datasets: Curate small-to-medium CERN-open datasets (simulated ttbar events, tracking ntuples) hosted on EOS or CVMFS. Include “toy” versions for quick iteration.
• Accessibility edge cases: Support CPU fallbacks, screen-reader-friendly notebooks, and asynchronous recordings for hybrid participants.

4. ENGAGEMENT, INCLUSIVITY, AND COMMUNITY BUILDING

• Interactive elements: Live coding, peer debugging, “GPU speed dating” with mentors, and a final demo/pitch session.
• Diversity angle: Actively invite underrepresented groups and feature lightning talks from junior researchers.
• Multi-angle value: Technical depth for experts plus “why GPUs matter for your physics” for newcomers.

5. POST-WORKSHOP FOLLOW-THROUGH: SUSTAIN MOMENTUM

• Materials & resources: All notebooks, recordings, and a “HEP GPU cookbook” repo.
• Ongoing support: Dedicated Mattermost/Slack channel (or HSF forum), monthly “GPU office hours,” and a 3-month “adoption challenge” with NVIDIA/CERN mentors.
• Impact measurement: Pre/post surveys, GitHub metrics, and a 6-month check-in on real workflows accelerated.

Potential challenges and mitigations:

• Varying expertise → Tiered tracks + pre-assessments.
• Resource contention → Batch scheduling + cloud fallback.
• Sustainability/energy → Include a session on profiling for efficiency.
• Long-term integration → Emphasize maintainable code and interoperability with existing C++/ROOT stacks.

By focusing on CERN-specific, hands-on, integrated content with strong pre/post support, this workshop can become a catalyst — seeding new working groups and positioning CERN as a leader in GPU-accelerated open science.

I’m happy to help refine the survey, draft the repo structure, or even co-moderate a session. Please forward any specific constraints (dates, venue capacity, budget) so we can iterate.

Looking forward to making this a standout event for the community!

Best regards,
Agamya Samuel