Solution Brief
One AI service. Data stays on campus.

Every lab.One AI service.

Campus-wide private AI on AMD Instinct™ MI350P.Every lab gets its own model on one LIQID UltraStack server, with research data kept inside your data center.

Download PDFTalk to an Expert
10
GPUs
MI350P on one host, room for 30
1.44
TB
HBM3E on one server
6
GPUs
of lab endpoints, one model each
0
hops
network hops between GPUs

The Pilot-to-Campus Dilemma

Keep private AI with one research center, or scale it to every lab without building a cluster.

Demand Pressure

Every lab wants AI

Departments ask for private model access at the same time.

Unsanctioned API use

Grant money and protected data drift to public APIs.

Infrastructure Pressure

GPUs locked in silos

Lab-owned servers sit idle while other labs wait.

Clusters add overhead

Multi-node builds need GPU networking, switches and staff.

Compliance Pressure

Protected data can’t leave

IRB protocols and data use agreements bar public AI APIs.

Findings must be traceable

Every coded theme has to link back to its source.

Share It. Assign It. Reassign It.

One 10-GPU LIQID UltraStack server runs a campus research AI service. Each lab or project gets a full GPU with its own model, and GPUs move to new projects each term or grant cycle.

  • Research data stays on campus
  • One model per GPU, no noisy neighbors
  • Fixed capacity, no per-token fees on grants
GPU allocation for a campus AI service with open-weight models, one model per GPU: 6 GPUs for lab endpoints, 2 for analysis, 2 for batch coding, on a LIQID GPU chassis with 10 AMD Instinct MI350P and 1.44 TB HBM3E, connected over PCIe Gen 5 to one dual-socket AMD EPYC 9005 host with one OS image.

Lab Endpoints · 6 GPUs

gpt-oss-120b, one full model per GPU, assigned to a lab or project.

~65 GB
fits one card
1,256
tok/s on 8 GPUs, 72B

Analysis · 2 GPUs

Llama 3.3 70B codes themes and summarizes, citing the source.

254
tok/s on 2 GPUs
~8
live researchers*

Batch Coding · 2 GPUs

Llama 3.1 8B labels survey and free-text responses at volume.

9,193
tok/s on 2 GPUs
9.5×
scaling, 1 to 10 GPUs
Measured: LIQID Phase-3 testing on 1 to 20 MI350P GPUs on one host. Named models are projected. *Illustrative; notes on page 2.
LIQID UltraStack with AMD MI350P for Higher Education

Proof: Measured on MI350P

No cluster, no idle silos: one host, ten GPUs, a model per lab, and room to grow to 30.

Result
What was measured
Why it matters here
~65 GB
gpt-oss-120b at MXFP4
Published model size: the full model fits on one 144 GB MI350P with room for cache
Six lab endpoints, no model split across cards
1,256 tok/s
on 8 GPUs
72B-class model, FP8, under interactive limits: P99 TTFT ≤ 2 s, ≥ 30 tok/s per user
Reference for interactive speed at campus scale
9.5× scaling
1 to 10 GPUs, 7B-class
4,653 tok/s on 1 GPU to 44,383 tok/s on 10, one replica per GPU
Every added GPU adds capacity for the next lab
What it means
Every lab gets private AI
One host, no GPU network to run
Model size is a published spec. 72B-class: FP8, one replica per GPU (254 tok/s on 2 GPUs, 552 on 4, 1,256 on 8). 7B-class: FP4, one replica per GPU, concurrency 128 (4,653 tok/s on 1 GPU, 9,193 on 2, 44,383 on 10). Different test conditions; don't compare them. *~8 = 254 ÷ 30 tok/s per user, illustrative. Source: LIQID Phase-3 testing. Named-model performance is projected from same-class results and has not been measured directly.

Grow on the Same Server: 10, 20 and 30 GPUs

Add GPUs and a new job, not a new cluster.

Start here
10
GPUs
1.44 TB HBM3E

Private AI service for many labs

One GPU per lab or project, reassigned each term or grant cycle.

Model layout

6 × gpt-oss-120b lab endpoints · 2 × Llama 3.3 70B analysis · 2 × Llama 3.1 8B batch coding

Built for

CIO / Director of Research Computing. Labs asking for AI; unsanctioned API use on grants.

20
GPUs
2.88 TB HBM3E

Research administration suite

Grant compliance review, export-control screening and proposal drafting.

Model layout

6 × 70B proposal and compliance review · 4 × 8B export-control screening · 10 × lab endpoints and analysis

Who buys, why now

VP Research / Sponsored Programs. Proposal volume; review backlog.

30
GPUs
4.32 TB HBM3E

University-wide private model hub

Research, administration and teaching on one host.

Model layout

14 × lab endpoints and analysis · 10 × research administration · 6 × gpt-oss-120b courses and coursework

Who buys, why now

CIO / Provost. Consolidating departmental AI; teaching demand.

One host with 30 GPUs serves research, administration and teaching. Starting with one research center? See the 4-GPU Higher Education partner brief.

What UltraStack Adds

  • One model per GPU: 144 GB HBM3E holds a full 70B model at FP8, or gpt-oss-120b.
  • No network between GPUs: peer-to-peer over PCIe Gen 5 to one EPYC host.
  • Air-cooled: about 22 kW for the full 30-GPU system.

Start Fixed, Grow on One Host

  • One host: one OS image, one license footprint, one thing to patch.
  • Room to grow: 10 to 20 to 30 GPUs on the same host.

What Is Proven Today

  • P2P bandwidth saturates the link; RCCL collectives pass, 1 to 10 GPUs.
  • ROCm integration complete, developed jointly with AMD.
  • 200+ production deployments, 85+ patents.
The Efficiency Case · Targets, not measurements*
75%
Up to 75% fewer servers
2.1×
Up to 2.1× more tokens per dollar
1.8×
Up to 1.8× better tokens per watt
50%
Up to 50% lower power

Bring Us Three Numbers

Two questions settle the size of the opportunity: how many labs are asking for AI today, and how many hold data they can't send to a public API?

Size Your CampusContact UsDownload PDF
1
Labs:

labs and active projects asking for AI.

2
Data:

projects with restricted data and their volume.

3
Users:

researchers working at peak.

Built for

Research computing centers · Human-subjects research · Campus-wide AI services

Proven on

AMD Instinct™ MI350P · AMD EPYC™ 9005 · ROCm™ · 200+ LIQID deployments

*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.

Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models. Outputs support, and do not replace, researcher review. Institutions remain responsible for IRB, data use agreement and export-control compliance.

Speak With An Expert