GPU Instances
Understanding Instance Selection
You need to choose a GPU instance that matches your computational requirements. NeevCloud offers various GPU configurations optimized for deep learning, AI training, and compute-intensive tasks.
Using Filters to Find Your Instance
You can narrow down your options using three key filters:
Region
Select a data center location close to you or your data source. Closer proximity reduces latency, which is critical when you're:
Transferring large datasets to your instance
Running real-time inference applications
Frequently accessing your instance via SSH or Jupyter
Collaborating with team members in specific geographic areas
Available regions typically include major data centers across North America, Europe, and Asia-Pacific.
VRAM
Choose the VRAM capacity your workload requires. Available options range from 4 GB to 128 GB.
Your VRAM requirement depends on:
Model size: Larger models require more VRAM to hold weights in memory
Batch size: Bigger batches need more VRAM for activations and gradients
Sequence length: For transformers, longer sequences consume more memory
Fine-tuning vs. inference: Training requires 3-4x more memory than inference
VRAM Guidelines:
4-8 GB: Small models, inference, lightweight training
16-24 GB: Medium models, fine-tuning smaller LLMs, computer vision
40-48 GB: Large models, full fine-tuning, multi-GPU workloads
80-128 GB: Very large models, research workloads, extensive batch processing
GPU Brand
Filter by hardware manufacturer. You'll find instances powered by NVIDIA, (AMD, and Intel coming soon) GPUs.
Brand Considerations:
NVIDIA: Widest framework support, best for PyTorch/TensorFlow, CUDA ecosystem
AMD: Good for specific workloads, ROCm support growing (coming soon)
Intel: Emerging option, strong for inference and certain HPC tasks (coming soon)
Different brands may offer specific optimizations for certain frameworks or workloads. Most templates are optimized for NVIDIA GPUs due to their dominance in the AI/ML space.
Important Considerations
Keep in mind that instance availability and pricing fluctuate based on your selected region and current demand. If your preferred configuration isn't available in one region, check alternative regions or adjust your specifications.
Availability factors:
Peak usage times (typically business hours in each region)
Seasonal demand (increased during conference deadlines, academic terms)
Specific GPU model popularity
Making Your Selection
After applying your filters, review the available instances and select the one that best matches your:
Computational requirements
Training: Higher VRAM, more compute power
Inference: Lower VRAM acceptable, focus on throughput
Research: Flexible configurations for experimentation
Budget constraints
Balance performance needs with hourly costs
Consider reserved instances for long-running jobs (more on this in the pricing section)
Performance expectations
GPU generation matters (newer = faster, more efficient)
Memory bandwidth affects data-intensive operations
Compute capability affects specific operations (like tensor cores)
Framework compatibility needs
Verify your chosen template supports your selected GPU
Check CUDA compute capability requirements for your code
Last updated