Research and education

The workload

Research and teaching compute is defined by its users rather than by one job. A single lab might run model training, inference for a demo, a simulation and a dozen student notebooks in the same week. The hardware has to be shared fairly, survive inexperienced users, be easy for a small IT team to keep running, and be bought on a grant or a budget cycle that does not move. Peak performance matters less than availability: a GPU a student can use today beats a faster GPU in a queue.

There are three recurring shapes. The principal investigator's workstation, which is also the lab's shared dev box. The departmental server, a few GPUs in a rack that a scheduler divides among users. And the institutional cluster, where a central team provides GPU nodes to the whole campus.

What it needs from the hardware

Memory per GPU, because research code is rarely optimized and a model that fits is a model that runs. Hardware partitioning where many small jobs share one card: RTX PRO 6000 Blackwell supports Multi-Instance GPU, splitting a single 96 GB card into up to four isolated instances, which turns one GPU into four student seats with no interference. Reliability features such as ECC memory, redundant power and remote management, because the person who fixes it may be an hour away. Enough CPU, host memory and local NVMe that data loading does not starve the GPUs when eight notebooks hit the same disk.

Software stacks matter more than usual: containers, a scheduler such as Slurm or Kubernetes, and a shared file system. We will install the base stack on request so that the system arrives ready for the scheduler.

Recommended systems

Workstation or small server

The AI developer workstation with one RTX PRO 6000 Blackwell is the research workstation that does everything: training small models, serving a demo, rendering figures. For a lab that wants more seats, the 4 GPU AI workstation with MIG enabled gives up to sixteen isolated instances from one quiet box under a bench.

PCIe GPU server

The departmental standard is a 4 GPU 4U PCIe server or an 8 GPU PCIe server with RTX PRO 6000 Blackwell Server Edition under a scheduler. Eight 96 GB GPUs with MIG serve a graduate program. For a smaller teaching lab, a 4 GPU 2U PCIe server with RTX PRO 6000 Blackwell Server Edition gives up to sixteen MIG seats in two rack units. Where ECC and MIG are not needed, the 4 GPU RTX 5090 AI server is the lowest cost per GPU for a teaching lab, a course inference endpoint or a student cluster, and for a single bench the dual RTX 5090 content creation workstation is the lowest cost per seat. GeForce cards are consumer products without data center driver licensing, so the lab confirms the intended use on the quotation.

HGX node or cluster

Where the institution trains large models or runs FP64 simulation at scale, the 8 GPU HGX B200 AI server is the node of choice, and several of them with an InfiniBand fabric form a campus cluster. We can quote nodes, fabric, storage and the acceptance testing as one line for a grant application.

Sizing notes

Count users rather than jobs: plan one MIG instance or one GPU per active user at peak, with 24 GB of GPU memory as the minimum useful instance for current models. Give shared servers 6 CPU threads and 24 GB of host memory per GPU for notebooks and data loading on top of the usual rules, and local NVMe of four times the GPU memory. Put a shared file system in from the start; a server with eight GPUs and one local disk becomes a storage problem within a term. Keep one spare GPU card in the department, because warranty replacement takes longer than a semester deadline allows.

Next step

Send us the number of users, the courses or projects, the budget line and the deadline for the purchase order, and we will size the system and quote it in the format your procurement office needs, within one business day. We are used to grant timelines, fiscal year ends and the questions purchasing departments ask, and we can provide sole source justification language where the specification supports it.