{"resourceId":"indiana-university-quartz-208-h100-gpus-queue-waits-20261019","versions":[{"version":"external-f75c3be19122a6c20be5653485eeb97a028b1f871f5866a8b86b9fdb4df18d86","resource":{"id":"indiana-university-quartz-208-h100-gpus-queue-waits-20261019","title":"Indiana University adds 208 NVIDIA H100 GPUs to its Quartz HPC system after 312 users faced waits of eight hours or more, and slices nodes to a quarter per job","organization":"Indiana University (UITS Research Technologies)","sector":"Public research university (research computing)","geography":"United States (Indiana)","publishedAt":"19 October 2026","publicationDate":"2026-10-19","eventDate":null,"sourceName":"Indiana University IT News","sourceLabel":"IU IT News, \"IU expands GPU computing capacity to accelerate research\", 19 October 2026","sourceUrl":"https://news.iu.edu/it/live/news/53945-iu-expands-gpu-computing-capacity-to-accelerate","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["infrastructure","operating-model","accessibility-workforce"],"finding":"IU IT News reported on 19 October 2026 that UITS Research Technologies added 52 Lenovo SD665-N nodes to the Quartz HPC system, each with two AMD EPYC 9454 CPUs, 1.5 TB of RAM and four NVIDIA H100 80GB GPUs, for 208 new GPUs. The stated demand case: in the prior six months Quartz and Big Red 200 served 1,189 researchers and students, nearly half using GPU partitions; 312 users waited eight hours or more for jobs to start, affecting over 34,000 jobs; one lab ran 7,308 AlphaFold3 jobs on only 48 H100s. Usage analysis showed most GPU jobs need one GPU, so most new nodes will be sliced to a quarter of a node per job to stop oversized requests locking out others. Expected reductions in wait times are aims, not measured results.","sledRelevance":"Gives a rare set of concrete demand numbers (users, wait times, jobs affected) behind a campus GPU purchase, and a specific allocation policy, node slicing based on usage analysis, that other research computing groups can compare against their own queues.","evidence":"Opened the IU IT News article dated 19 October 2026 and read the full text, including usage statistics, node specifications, the slicing policy and the researcher quote.","architectureImplications":"Adds H100 nodes within an existing cluster rather than a new system; per-job slicing to one GPU and a quarter of a node assumes most workloads fit one 80GB GPU, which usage data supported.","governanceImplications":"Allocation policy is set from historical usage analysis, an approach that is auditable and adjustable; the article does not describe exceptions for multi-GPU jobs.","securityPrivacyImplications":"Not addressed in the article.","caveats":"IU's own IT news item. Demand statistics are self-reported without methodology, and the article gives no cost, funding source or post-deployment wait-time data. A researcher quote says \"200 new H100\" GPUs while the article body says 208.","streamIds":["research"],"roles":{"sales":"Lead with queue evidence, not GPU counts. IU justified 208 new H100s with concrete numbers: 312 users waited eight hours or more and over 34,000 jobs were affected. Research computing directors and CIOs need that kind of evidence to win capacity funding. A bounded engagement analyses scheduler logs and models how capacity and allocation policy would change waits. Interpretation: IU reports demand and plans, not results, so it cannot support claims that buying GPUs cuts waits by a set amount.","engineering":"Treat allocation policy as part of the design. IU found most GPU jobs need one GPU and will slice new four-GPU nodes to a quarter per job. Pre-sales should analyse scheduler logs for per-job GPU and memory use, check whether multi-GPU jobs need a separate partition, and confirm memory headroom on 80GB cards. Interpretation; the article gives no scheduler, network or storage detail.","delivery":"Frame delivery as running a demand-driven GPU service. Work covers usage analysis, partition and slicing policy, user communication and monitoring of wait times and utilisation after the expansion, with a route for multi-GPU jobs. Central research computing owns it with faculty governance for allocation rules. Interpretation: acceptance should rest on measured wait times and utilisation before and after, because the source reports pre-expansion demand only."},"roleDetails":{"sales":{"customerProblem":"Research computing groups face long GPU queues and need evidence to justify and allocate new capacity.","stakeholders":["Research computing director","Chief information officer","Vice president for research","Faculty HPC advisory committee"],"discoveryQuestions":["What share of GPU jobs use only one GPU?","How many users wait more than eight hours for jobs to start?","Which workloads need newer GPU generations?"],"valueHypothesis":"Scheduler-log analysis may let an institution right-size a GPU purchase and allocation policy before spending capital.","potentialEngagement":"A three- to four-week scheduler-log analysis and capacity model with recommended partition and slicing policies.","unsupportedClaims":["A specific reduction in wait times","Cost-effectiveness of the purchase","Research output gains"]},"engineering":{"fit":"Fits existing clusters with measurable queue pressure from mostly single-GPU workloads; less fit for large multi-GPU training.","architectureIntegration":"New GPU nodes added to an existing scheduler with partitions that slice nodes per job and a route for multi-GPU work.","prerequisites":["Scheduler logs with per-job GPU and memory use","Facility power and cooling for new nodes","Partition and fair-share policy"],"constraints":["Slicing limits large multi-GPU jobs","Funding and cost not disclosed","Software requiring newer GPU generations"],"security":"Standard HPC account and data controls; confirm GPU sharing does not expose memory between jobs under the chosen slicing method.","proofOfValue":["Compare median and 90th-percentile wait times before and after expansion","Measure GPU utilisation on sliced nodes","Track multi-GPU job wait times separately"]},"delivery":{"work":"Analyse usage, set partition and slicing policy, communicate changes to users and monitor waits and utilisation.","dependencies":["Node installation and acceptance testing","Faculty agreement on allocation rules","Monitoring dashboards"],"ownership":"Central research computing with faculty HPC governance.","skillsAdoption":"User guidance on requesting right-sized GPU jobs; workshops for labs moving from older GPUs.","governanceCheckpoints":["Allocation policy approval before launch","Three-month wait-time review","Annual capacity review"],"acceptanceCriteria":["Proposal: median GPU job wait below an agreed target three months after launch","Proposal: share of users waiting over eight hours reported quarterly","Proposal: sliced-node utilisation above an agreed threshold"],"risks":["Demand growth outpacing new capacity","Multi-GPU users disadvantaged by slicing","No post-launch measurement"]}},"retrievedAt":"2026-10-20T22:05:46-05:00","enrichedAt":"2026-10-20T22:05:46-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Expanded access is framed for both researchers and students; no training or support staffing changes are described.","procurementImplications":"Incremental node purchase into an existing system; cost and vendor terms are not given.","operatingModelImplications":"Central research technologies run allocation using usage data, which makes wait-time and utilisation monitoring an ongoing operational duty.","sourceVerification":{"openedUrl":"https://news.iu.edu/it/live/news/53945-iu-expands-gpu-computing-capacity-to-accelerate","referenceExcerpt":"312 users faced wait times of eight hours or more for their computing job to start running","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}