Requesting access¶
Who can use FLAME?¶
FLAME is targeted research infrastructure for AI/ML and data-intensive science and engineering. It's currently available to Drexel research groups doing work in those areas. In particular, FLAME is not currently available for classroom use.
FLAME cannot host regulated or restricted data
FLAME is not an approved environment for:
- Data under regulatory protection (e.g. HIPAA, FERPA, or similar laws)
- Data under an IRB protocol with specific handling requirements
- Data under a Data Management Plan, Data Use Agreement, or other contractual restriction imposing specific security, access, or handling requirements
Research involving any such data will not be approved. Please confirm your data is free of these restrictions before applying.
How to request access¶
All access to FLAME is granted through allocations, which a PI requests on behalf of their research group using the application form here:
If you're a PI, click the button above to start your request. If you're a student or research staff, please send your PI this page and ask them to fill out the form.
How are requests reviewed?¶
Applications are reviewed by a steering committee, which consists of the PIs on the MRI that funded the cluster and URCF technical staff. The Committee considers the following when evaluating requests for FLAME access:
- FLAME alignment. The research is in AI/ML or data-intensive science and engineering, and fits FLAME's capabilities (available GPUs and software environment).
- Readiness. Your research is developed enough to quickly make productive use of the allocation, and the group members who'll use FLAME have demonstrated capability with GPU computing.
- Track record. Publication history, current or prior funding, or other concrete indicators of research output.
- FLAME's available capacity at the time of application.
An application may:
- approved as submitted
- modified (e.g. GPU count or duration adjusted) and approved
- deferred until FLAME has more available capacity
- declined if the proposed research is not suitable for FLAME
What is an allocation?¶
An allocation grants your research group access to specific GPUs for a time-bounded period. For example, "two H200s for 30 days" or "one GH200 for three months." Allocations are granted to a research group and shared between all members of that group.
For more details on how allocations work, see the Allocations reference page.
Which type of allocation should I request?¶
There are two types of allocation, and they tradeoff dependability against ease of approval.
A standard allocation reserves a specific number of GPUs for your group. That capacity is guaranteed: you can run on those GPUs the moment you need them. This dependability comes at a cost, though — guaranteed capacity is finite, so standard allocations are reviewed against what's currently unallocated, and a request may be modified or deferred if FLAME is near capacity.
A borrowing-only allocation reserves nothing. Instead, it lets your group use GPUs that standard holders aren't using at the moment. When the cluster is busy, your workloads may queue until GPUs free up, and once running they can be evicted so a standard holder can reclaim those GPUs. Borrowing-only workloads have to be patient and interruptible — they run when there's room and yield when there isn't. The advantage is that borrowing-only allocations tie up no guaranteed capacity, so the Steering Committee can approve them much more freely. They're good for exploratory or bursty work, smaller or occasional needs, or just trying FLAME out before requesting a standard allocation.
| Standard allocation | Borrowing-only allocation | |
|---|---|---|
| Guaranteed GPUs | Yes | No |
| Interruptions | Possible | |
| Ease of approval | Harder | Easier |
| Best for | Dependable, on-demand work | Exploratory or bursty work; getting started |
If your work needs capacity you can count on being available at any time, apply for a standard allocation. If it's flexible or you're still getting up to speed, a borrowing-only allocation is a better fit. You can request either through the same access request form.