Academic GPU Computing: For Research Groups
Dedicated GPU computing capacity for research groups, fundable through TÜBİTAK and EU Horizon project budgets.
Research groups face a recurring GPU access problem: either there is no hardware, the shared cluster queue takes weeks, or the institution’s procurement process does not align with the pace of research. Yet for training deep learning models, running large-scale molecular dynamics, or completing genomic pipelines, a GPU is not an optional accessory — it is the research itself.
Mevasis offers research groups and higher education institutions dedicated, rented GPU capacity for the duration of their projects. Hardware is located in Turkey; data does not leave the country; and the contract structure is compatible with TÜBİTAK and EU Horizon budget categories.
Why GPU, Why Now?
The transformation in computational research over the last five years has moved in one direction: from CPU to GPU. This shift is not limited to deep learning. Molecular dynamics, genomic variant analysis, radiation transfer simulation, and image-based scientific computing have largely migrated to GPU-accelerated versions.
A concrete example: a 100 ns GROMACS simulation (100,000 atoms) takes approximately 72 hours on 64 CPU cores, but only 4–5 hours on a single NVIDIA A100 80 GB. Running four A100s in parallel completes the same job in 1–1.5 hours. From a research cycle perspective, this means compressing a multi-week experimental phase into a single day.
In academic settings, however, sustained access to this hardware is difficult. Institutional GPU clusters are often insufficient or divided among multiple departments. National computing infrastructure (TÜBİTAK ULAKBİM) is a valuable resource, but its application and approval process is incompatible with urgent experimental needs. Cloud GPUs raise serious questions around data sovereignty, KVKK compliance, and cost predictability in a research environment.
Typical Academic GPU Workloads
The table below summarises common GPU workloads observed in research groups at Turkish universities, along with relevant software tools.
| Research Area | Core Workload | Tools Used | Recommended GPU |
|---|---|---|---|
| Machine Learning / Deep Learning | Model training, fine-tuning, hyperparameter search | PyTorch, TensorFlow, JAX, Hugging Face | A100 80 GB / H100 |
| Molecular Dynamics | Protein-ligand simulation, membrane dynamics | GROMACS, NAMD, AMBER, OpenMM | A100 / L40S |
| Genomics | NGS pipeline acceleration, variant analysis | NVIDIA Parabricks, BWA-MEM2, DeepVariant | A100 / RTX 4090 |
| Protein Structure Prediction | Single-chain and multi-chain structure prediction | AlphaFold 2 / 3, RoseTTAFold, ESMFold | A100 80 GB |
| Computer Vision | Microscopy image analysis, segmentation | PyTorch, MONAI, cellpose | L40S / RTX 4090 |
| Physics Simulation | Particle physics, fluid dynamics (GPU CFD) | Geant4 GPU, AmgX, PhysX | L40S / A100 |
| Large Language Models (LLM) | Inference, RAG, fine-tuning | vLLM, llama.cpp, Ollama | H100 / A100 |
| Computational Chemistry | GPU-accelerated DFT, docking | ORCA GPU, AutoDock-GPU, GNINA | RTX 4090 / A100 |
Note: GPU memory capacity is a critical selection criterion. AlphaFold 3 requires 80 GB VRAM for large proteins. LLM fine-tuning may require multi-GPU (NVLink-connected) configurations depending on model size.
Sector Use Cases
AI Research Groups
The most common demand in computer engineering and AI laboratories is dedicated GPU capacity per project. Multiple PhD students training different models simultaneously causes conflicts in shared environments. With SLURM-based job queuing, GPU quotas can be assigned per user or subgroup, with priority policies set according to the group’s own rules.
For model training workflows, tracking tools such as Weights & Biases, MLflow, or TensorBoard integrate seamlessly with the infrastructure. Jupyter Lab and VS Code Remote-SSH access are provided as standard.
Biomedical and Genomics Research
Under KVKK, processing personal health data on overseas servers creates legal risk. NVIDIA Parabricks accelerates the standard GATK Best Practices pipeline for NGS analysis by approximately 50×; achieving this advantage on Turkey-located GPUs meets data sovereignty requirements without compromise.
AlphaFold and RoseTTAFold setup (including ~2.2 TB database requirements) is provided turnkey — the research group spends its time on research, not installation.
TÜBİTAK and EU Horizon Projects
In TÜBİTAK 1001, 1003, or ARDEB-funded projects, computing infrastructure costs can be budgeted under the “service procurement” line item. In EU Horizon Europe projects, GPU rental costs are accepted under “other direct costs” or “infrastructure usage fees.”
Mevasis offers a contract structure aligned with the project duration (typically 12–36 months). When the project closes, the infrastructure obligation ends — no fixed hardware investment required. Invoice formats compatible with TÜBİTAK budget reporting are provided.
Multi-Group / Faculty-Level Shared Access
For cluster configurations shared by multiple research groups, SLURM’s fair-share algorithm allows GPU-hour quotas to be assigned per group. Billing can be issued separately for each group based on consumption. This model is particularly flexible for newly established groups or those requiring capacity during specific project periods.
Reference GPU Cluster Configuration
Mevasis reference architecture for a medium-sized research group or multi-group shared environment:
Academic GPU Cluster — Reference Configuration
Login / JupyterHub Node (1–2 units)
├── GPU Compute Nodes — Training (4–8 units)
│ ├── 2× Intel Xeon Gold 6438M (32-core)
│ ├── 512 GB DDR5 ECC
│ └── 4× NVIDIA A100 80 GB SXM (NVLink-connected)
│ Total: 16–32 GPUs / 1,280–2,560 GB VRAM
│
├── GPU Compute Nodes — Inference / Mid-Scale (2–4 units)
│ ├── 2× AMD EPYC 9354 (32-core)
│ ├── 256 GB DDR5 ECC
│ └── 4× NVIDIA L40S 48 GB
│ Total: 8–16 GPUs
│
├── CPU High-Memory Nodes (2 units, for genomics / large MD)
│ └── 2× AMD EPYC 9654, 1.5 TB DDR5
│
└── Storage
├── NVMe Scratch: BeeGFS, 4–10 GB/s, 100 TB+
└── Project Data: Redundant SAS/NL-SAS, scalable
Network:
├── Between GPU nodes: InfiniBand HDR200 (NVLink workloads)
└── General access: 25 GbE
Software Layer:
├── Job scheduler: SLURM + GPU GRES resource management
├── Container: Singularity / Apptainer (Jupyter, PyTorch, TF images)
├── Module management: Lmod + EasyBuild
└── Monitoring: DCGM (GPU metrics) + Prometheus + Grafana
Data security: Hardware located in Turkey is the default safe choice for KVKK compliance. For research involving clinical, genomic, or personal data, overseas cloud alternatives require additional legal review.
License Compliance
Open-source tools (GROMACS, NAMD, PyTorch, TensorFlow, Parabricks Academic, AlphaFold, OpenMM, LAMMPS, BWA-MEM2, STAR) can run on Mevasis infrastructure without additional licensing costs.
For commercial software (MATLAB, ANSYS Mechanical / Fluent, Gaussian, AMBER), the institution’s existing license agreement should be verified. HPC-packaged or token-based licenses are generally compatible with Mevasis infrastructure; license server integration is provided by Mevasis.
NVIDIA NGC images (including PyTorch, TensorFlow, RAPIDS, and CUDA tools) can be used directly in the Mevasis environment; container support is standard.
Mevasis Academic GPU Services
Service scope offered to research groups and higher education institutions:
- Dedicated GPU rental: GPU capacity assigned to a group or project, billed monthly or per project duration. See HPC Rental
- Turnkey software installation: AlphaFold, GROMACS, NAMD, Parabricks, PyTorch / TF environment, JupyterHub setup and configuration. See HPC Support & Consulting
- GPU workload profiling: Analysis of your existing code and SLURM history to determine the optimal GPU type and count.
- Multi-group management: SLURM fair-share, GPU quotas, and department-level billing configuration.
- Data security consulting: KVKK compliance, data classification, and research data retention policies.
Frequently Asked Questions
Why Mevasis over cloud GPU for research? Cloud GPUs create two fundamental problems in research projects: cost unpredictability and data sovereignty. Spot instances can be interrupted during workloads; on-demand pricing makes monthly budget planning difficult. Clinical and genomic data covered by KVKK cannot be processed on overseas cloud providers. Mevasis eliminates both problems with a fixed monthly rate and Turkey-based location.
Can GPU rental be funded from a TÜBİTAK project budget? Yes. GPU computing rental costs can be budgeted under “service procurement” or “computer usage fees” in TÜBİTAK 1001/1003 projects. Contract and invoice structures are tailored to the specific project during the budget planning phase.
Can multiple research groups share the same cluster? Yes. SLURM assigns GPU-hour quotas and priority scores (fair-share) to each group. The workloads of different groups do not interfere with each other; consumption-based invoices can be issued separately for each group.
Is AlphaFold or Parabricks installation included? Yes. Turnkey installation, version tracking, and database updates (approximately 2.2 TB for AlphaFold) are covered by Mevasis support. The aim is for research groups to focus on research rather than managing technical infrastructure.
Can I run my own CUDA code on this infrastructure? Yes. CUDA 12.x and related tools (cuDNN, NCCL, cuBLAS) come pre-installed. You can run custom container images with Singularity/Apptainer, and MPI + CUDA hybrid workloads are fully supported.
To discuss your research group’s GPU needs, contact us. We will provide a configuration recommendation suited to your project budget and workload profile.