From the Dell edition of October 31, 2026
University of Washington Tillicum: 24 Dell XE9680 servers with 192 H200 GPUs delivered on a narrow public budget, per partner case study; TOP500 record shows rank slipping from 185 to 207
University of Washington (with Dell partner The Redesign Group) · Higher education / research computing · Seattle, Washington, United States
- Publisher
- The Redesign Group insights (Danny James)
- Original publication
- 2026-10-21
- Source retrieved
- 2026-11-01
- Event date
- 2025-10-15
What happened
A case study published Oct. 21, 2026 by Dell partner The Redesign Group describes how the University of Washington built Tillicum, its GPU research platform, under a narrow public-university budget. Phase 1 (about 200 GPUs, a bridge to a planned 1,000+ GPU next-generation phase) comprised 24 Dell XE9680 8-way H200 servers, 3 PB of all-flash storage, NVIDIA NDR InfiniBand, Dell factory rack integration and white-glove installation; Redesign says it worked with Dell, NVIDIA and DataCore and used its partner status to get within budget. It claims research met utilization goals and that UW 'maintained and even moved up' to TOP500 #185 and Green500 #51. UW's own documentation confirms 24 XE9680 servers, 192 H200 GPUs and ~3 PB flash, with general availability on Oct. 15, 2025. The TOP500 record shows #185 at debut (Nov. 2025) and #207 in June 2026, so the 'moved up' claim is not supported.
Why it matters
One of few documented public-university AI clusters built on Dell servers with a stated budget constraint, phased roadmap and partner-led procurement; relevant to research universities and state systems planning shared GPU capacity. It also shows how partner marketing can overstate outcomes that public records can check.
Evidence and measured results
Partner marketing case study; no cost, utilization numbers, user counts or research outputs are disclosed. Independent corroboration: TOP500 lists the Dell-built system at 8.85 PFlop/s Rmax and 144 kW, rank 185 (11/2025) and 207 (06/2026); UW Research Computing documentation confirms the hardware and an Oct. 15, 2025 general-availability date.
Limitations and uncertainty
Partner-authored marketing; the deployment dates to 2025 and the case study is new (Oct. 21, 2026), which is why it is included now. Utilization and budget outcomes are unquantified. The TOP500 'moved up' claim conflicts with TOP500 records; Green500 #51 was not independently checked in this run.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-01; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Public research universities need shared GPU capacity for AI research quickly, within constrained budgets and existing data center limits.
- Stakeholders
- VP for research / research computing director
- Central IT/CIO
- Procurement office
- Faculty governance for allocation
- Facilities
- Discovery questions
- Is a bridge phase needed before a larger liquid-cooled system?
- What power and cooling headroom exists in current data centers?
- How will GPU time be allocated and charged across departments?
- Which contract vehicle gives the best verifiable price?
- Value hypothesis
- A phased, factory-integrated Dell GPU cluster may let a university add research AI capacity quickly; cost and utilization benefits are not disclosed.
- Potential engagement
- A bounded capacity-planning and architecture review for a phase-one research GPU cluster.
- Unsupported claims
- That the system moved up the TOP500
- Any specific utilization rate or cost savings
- That partner pricing will be available to other institutions
Pre-sales engineering
Role takeaway
- Fit
- Good fit for research institutions with air-cooled data center space and a near-term GPU gap.
- Architecture and integration
- Dell XE9680 H200 nodes with NDR InfiniBand and shared flash storage under a Slurm-style scheduler with container support.
- Prerequisites
- Data center power and cooling capacity
- Allocation and cost-recovery policy
- Storage sized for AI datasets
- Constraints
- 144 kW Linpack power draw for the cluster
- Next phase may require liquid cooling or colocation
- Restricted data requires separate environments
- Security
- Not described; require separate controls or environments for regulated research data.
- Proof of value
- Benchmark representative AI training and inference jobs
- Measure queue times and utilization during early access
- Verify vendor benchmark claims against TOP500 or local runs
Delivery
Role takeaway
- Work
- Install, integrate and open a shared GPU cluster with early access before general availability.
- Dependencies
- Factory rack integration
- Facility readiness
- Scheduler and storage configuration
- Ownership
- Central research computing (UW Research Computing in this case).
- Skills and adoption
- Researchers need onboarding on scheduler, containers and GPU programming; training offered as an option.
- Governance checkpoints
- Facility readiness sign-off
- Allocation policy approval
- Early-access review before general availability
- Acceptance criteria
- Proposed: target GPU utilization and queue-time metrics defined and reported monthly
- Proposed: early-access users validate key workloads
- Proposed: phase-two design documented before phase-one saturation
- Risks
- Overstated vendor or partner outcome claims
- Facility limits for later phases
- Underused capacity without allocation policy
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?
Air-cooled 8-GPU H200 nodes with InfiniBand and shared flash as a bridge phase, with liquid-cooled Blackwell-class systems in a colocation or modular data center deferred to a later phase; factory rack integration shortened install.
Governance
Who approves, reviews and stays accountable for outcomes?
Partner discounts and escalations shaped what a public university could afford; institutions should document how partner pricing was obtained and verify outcome claims against public benchmarks before reuse.
Security and privacy
What data, permissions and controls need testing?
Not addressed; research clusters handling medical or restricted data need separate controls (UW runs a distinct restricted-access computing service).
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Targets researchers across all UW departments and teaching use; the case study lists end-user training as an optional service.
Procurement
What should contracts, pricing and exit terms secure?
Shows the role of a reseller's top-tier Dell partner status and vendor escalation in fitting a public budget; buyers should compare against cooperative contract pricing.
Operating model
Which teams own the service once it runs?
Operated by UW central research computing with a phased capacity plan; Day 2 management and build-operate-transfer options were offered by the partner.
Publication history
- 2026-10-31Dell · Issue 033 resources
Stable resource ID: uw-tillicum-dell-xe9680-h200-cluster-redesign-case-study-20261021