About
The Model
- Visible collaboration spaces for researchers, students and industry partners
- Demonstration and workshop environments to support engagement and translation
- Shared virtual AI infrastructure accessible across faculties and campuses
- Scalable virtual environments with a lower marginal cost to expand over time
- An integrated online community space for researchers to share experiences, expertise, and best practices as they learn to leverage high-memory resources to accelerate their research
Access Model
Access Model | Monash AI Co-Lab (Shared, On-Campus) | Dedicated Device | Remote Access (Digital Capability) |
|---|---|---|---|
| Overview | Shared, on-campus Spark workstations within Monash AI Co-Lab spaces | Individual Spark workstation purchased and owned by faculty/researcher | Remote access to Spark workstation (accessible by multiple users) |
| Which setup fits your research? | Best for fast access, support, and collaboration across larger research groups, student projects | Best for unlimited usage (but higher cost and effort associated with device setup) | Best long-term scalable model, but not available immediately |
| Access Speed | Fastest access (devices centrally provisioned) | Slower due to procurement and setup lead times | Not yet available |
| Device Cost | Subsidised | ~$8–9k per device + accessories (faculty/researcher-funded) | ~$8–9k per device + accessories (faculty/researcher-funded), costs associated with data storage |
| Usage Model | Shared devices (booking via Skedda) | Dedicated, unrestricted usage | Fully remote access |
| Support | eResearch (community of practice - shared knowledge, peer support) eSolutions (user access, OS support) | No managed support (device delivered unmanaged) | Expected to be centrally supported (TBC) |
| Collaboration | Strong – co-located, shared spaces, peer learning | Limited – individual device use for 1:1 usage | Remote collaboration only |
| Data | Data classified as Public or Restricted only. No storage or retention of data on these devices; data is retained for the session only and then removed | Data classified as Public or Restricted only. Research use involving Sensitive or Very Sensitive data will need to undertake an ISRA for advice | Data classified as Public or Restricted only. No storage or retention of data on these devices; data is retained for the session only and then removed |
Setup Effort | Managed Centrally via eSolutions | To be managed by the Faculty | Managed Centrally |
| Flexibility | On-campus access only | Full control and flexibility - assessed case by case basis | High (remote, cross-campus access) |
| Compliance | Managed within a university environment | Requires ISRA / Cyber assessment | Managed within a university environment |
| Limitations | Shared access - potential availability constraints if demand increases. Useable for interactive development - not available for long-running processes | Cost, setup burden, and ongoing costs associated with data storage | Under development - 6+ months required for development. |
Technology Enabler
The AI Co-Lab will leverage DELL GB10 as a core enabling technology. These compact, Linux-based AI compute devices are powered by the NVIDIA Grace Blackwell Superchip, providing 128GB of unified memory and supporting large AI models.
By providing dedicated access to large-memory machines within our physical spaces, we enable researchers to test and develop complex AI workloads that exceed the limits of standard hardware. Crucially, the DELL GB10 serves as a vital stepping stone, allowing researchers to prototype on the same architecture that powers MAVERIC, our large-scale NVL72 system, ensuring an easy transition from development to massive-scale production.
What the devices can do
The DELL GB10 are a desk-side AI appliance, not a general-purpose PC or workstation refresh. They are purpose built for AI development, prototyping, fine-tuning, RAG, multimodal and data science workloads. These machine bring a data centre-class AI performance (up to ~1000 FP4 TOPS, ~200B parameter models) to Monash.
The GB10s are ideal for:
- prototyping architectures
- fine-tuning on Monash datasets
- local inference/RAG on lab corpora
- data science
- feature engineering (RAPIDS, PANDAS, Polars)
- handling 80-90% of day to day research workloads at lower cost