From the NVIDIA edition of November 4, 2026
CUDA 13.5 makes Vera Rubin support generally available and adds MPS v3 GPU-sharing controls: weighted timeslicing, SM borrowing/lending and cgroup-based routing
NVIDIA · GPU software platform · Global
- Publisher
- NVIDIA Technical Blog
- Original publication
- 2026-11-04
- Source retrieved
- 2026-11-05
- Event date
- 2026-11-04
What happened
CUDA 13.5 moves Vera Rubin support from early developer access to general availability and extends Windows on Arm support to DGX Station for Windows. Multi-Process Service v3 adds weighted timeslicing, SM partition load balancing with borrowing and lending, cgroup v2 subtree-to-namespace routing, per-namespace SM partition assignments, and a polling-based fallback for memory accounting when cgroup notifications are missing or unreliable. Other changes cover CUDA Tile multi-GPU communication, cuda.core 1.2 and CCCL 3.5, including reproducible floating-point scans.
Why it matters
Shared campus and state GPU clusters often split GPUs among many users. MPS v3 gives administrators finer controls over that sharing, following the CUDA 13.4 Kubernetes memory enforcement covered on Oct 27. Vera Rubin GA matters to institutions planning next-generation purchases.
Evidence and measured results
Vendor release notes; no performance or fairness measurements for the new MPS features.
Limitations and uncertainty
Vendor announcement; feature claims are untested here; borrowing behaviour under contention is not quantified.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-05; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Shared GPU clusters struggle to allocate capacity fairly among many users.
- Stakeholders
- Research computing director
- HPC administrators
- Faculty user committee
- Discovery questions
- How are GPUs shared today?
- Are users waiting while GPUs sit partly idle?
- Is a Vera Rubin purchase being planned?
- Value hypothesis
- Finer sharing controls may raise utilisation on shared GPUs; gains must be measured locally.
- Potential engagement
- A sharing-policy pilot on one partition after a CUDA 13.5 upgrade.
- Unsupported claims
- Any utilisation gain
- Tenant security isolation via MPS
Pre-sales engineering
Role takeaway
- Fit
- Fits multi-user GPU clusters using MPS.
- Architecture and integration
- MPS v3 namespaces mapped from cgroup v2 subtrees; per-namespace SM partitions.
- Prerequisites
- CUDA 13.5 toolkit and drivers
- cgroup v2
- Launcher integration
- Constraints
- Feature behaviour under contention unquantified
- Driver/toolkit upgrade testing needed
- Security
- MPS is a sharing mechanism, not presented as an isolation boundary.
- Proof of value
- Measure queue wait and throughput before and after on one partition
- Test polling memory accounting overhead
Delivery
Role takeaway
- Work
- Introduce governed GPU-sharing policy.
- Dependencies
- CUDA 13.5 upgrade
- Scheduler configuration
- Ownership
- HPC operations manager.
- Skills and adoption
- Administrator training on MPS v3 and cgroups.
- Governance checkpoints
- Policy owner named
- User notice before weight changes
- Acceptance criteria
- Proposed: documented sharing policy
- Proposed: before/after utilisation report
- Risks
- Unfair allocation
- Upgrade regressions
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Sharing policy can follow the launcher's cgroup groupings, which ties scheduler configuration to GPU partitioning.
Governance
Who approves, reviews and stays accountable for outcomes?
Weighted timeslicing turns GPU-share priorities into an explicit policy that someone must set.
Security and privacy
What data, permissions and controls need testing?
Not addressed; MPS sharing is not presented as a security isolation boundary.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Administrators need MPS and cgroup skills to use the new controls.
Procurement
What should contracts, pricing and exit terms secure?
Vera Rubin GA support in CUDA reduces one software-readiness risk for buyers; hardware availability is separate.
Operating model
Which teams own the service once it runs?
Allocation policy for shared GPUs becomes more configurable and needs an owner.
Publication history
- 2026-11-04NVIDIA · Issue 605 resources
Stable resource ID: nvidia-cuda-13-5-vera-rubin-ga-mps-v3-sharing-controls-20261104