Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 24, 2026

Standards or public-body guidanceCautionaryUndated source

GPU Operator 26.7.1: CUDA memory limits, driver 615.71.09, H100 MIG upgrade known issue

NVIDIA · Kubernetes GPU infrastructure · Global

NVIDIA
Publisher
NVIDIA GPU Operator Docs
Original publication
GPU Operator 26.7.1 release notes (page undated; inspected 2026-09-24)
Source retrieved
2026-09-25
Read original source

What happened

GPU Operator 26.7.1 release notes add soft/hard per-GPU CUDA memory limits via NVIDIA_GPU_MEMORY_REQUEST and NVIDIA_GPU_MEMORY_LIMIT (MiB; requires R615+ driver), support for Data Center driver 615.71.09, Confidential Computing Manager v0.4.4, and operand wait-for-driver-validation fixes. A known issue warns that on H100 nodes, upgrading drivers from 595.91.07 to 615.71.09 without reboot or GPU reset can break subsequent MIG reconfiguration (NVML ERROR_UNKNOWN / Xid 119), with documented reboot or nvidia-smi -r workarounds. 26.7.0 previously introduced GPUCluster DRA management mutually exclusive with ClusterPolicy. Newly relevant beside NVCRE blogs that list GPU Operator as a prerequisite.

Why it matters

Campus/agency Kubernetes GPU fleets need memory-limit controls and must plan H100 MIG driver upgrade downtime—operational constraint evidence, not a SLED outcome study.

Evidence and measured results

Primary release notes detailing 26.7.1 features/fixes and H100 MIG known issue steps.

Limitations and uncertainty

Vendor docs; publicationDate not stated on page; behavior depends on cluster runtime/CDI configuration.

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-25; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.

Sales

Role takeaway
Customer problem
Multi-tenant GPU clusters need memory controls and safe driver upgrades.
Stakeholders
  • platform SRE
  • research computing
  • security
Discovery questions
  • H100 MIG in use?
  • Driver branch today?
  • Need per-GPU memory caps?
Value hypothesis
26.7.1 controls reduce neighbor risk if upgrade runbooks include MIG checks.
Potential engagement
Upgrade readiness review for one H100 pool.
Unsupported claims
  • Memory limits equal confidential computing
  • DRA GPUCluster is always preferred

Pre-sales engineering

Role takeaway
Fit
Operational guidance for 26.7.x fleets.
Architecture and integration
Operator-managed drivers/operands; exclusive ClusterPolicy vs GPUCluster.
Prerequisites
  • R615+ for memory limits
  • CDI-compatible runtime for some DRA paths
Constraints
  • H100 MIG upgrade known issue
  • Mutual exclusion of management modes
Security
Memory caps are isolation aids, not TEEs; CC Manager is separate.
Proof of value
  • MIG reconfig after upgrade
  • Memory limit rejected oversize alloc
  • Operands healthy

Delivery

Role takeaway
Work
Apply 26.7.1; execute H100 upgrade SOP; validate MIG and memory limits.
Dependencies
  • node maintenance windows
  • workload owners
Ownership
Kubernetes platform team.
Skills and adoption
Train on nvidia-smi reset and MIG state labels.
Governance checkpoints
  • CAB for driver upgrade
  • Post-change MIG test
  • Rollback driver pin
Acceptance criteria
  • No Xid 119 on MIG ops
  • Memory envs honored
  • ClusterPolicy ready
Risks
  • Skipped reboot
  • Mixed MIG/full-GPU scheduler issues on older drivers

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?

Enable memory request/limit env only with R615+; treat GPUCluster vs ClusterPolicy as exclusive modes; schedule MIG-sensitive upgrades with reset windows.

Governance

Who approves, reviews and stays accountable for outcomes?

Change control for driver upgrades must include MIG validation gates on H100.

Security and privacy

What data, permissions and controls need testing?

CC Manager bump is incremental; memory limits aid noisy-neighbor control but are not a TEE.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Ops procedure complexity; limited end-user accessibility impact.

Procurement

What should contracts, pricing and exit terms secure?

Specify supported Operator/driver pairs and maintenance windows in contracts.

Operating model

Which teams own the service once it runs?

Driver/MIG changes need formal change control tied to AI capacity calendars.

Publication history

  1. 2026-09-24NVIDIA · Issue 198 resources
Read preserved resource versions (JSON)

Stable resource ID: nvidia-gpu-operator-26-7-1-notes-2026