Skip to main content

​
Shared Network Bandwidth

GPUHub uses a shared bandwidth model for GPU instances. Network bandwidth and traffic are included with the instance and are not billed separately.
Network performance varies by region and host. Each region and host may have its own bandwidth capacity and transmission limits. The bandwidth available to an individual instance is not guaranteed and should not be interpreted as dedicated bandwidth. The total upstream bandwidth available in a region may be approximately 1–2 Gbps, but this capacity is shared among multiple hosts and instances. Actual upload and download speeds may vary depending on the region, host, concurrent network usage, destination server, network route, and time of day.

​
Network Performance During Peak Hours

During periods of high network activity, large dataset transfers by other users may reduce SSH responsiveness and file transfer speeds. This affects network-related operations only and does not directly affect local GPU training or computation performance.

​
Downloading Large Models or Datasets

Non-GPU Mode provides only limited CPU and memory resources and is intended primarily for tasks such as downloading models or datasets and preparing files, rather than compute-intensive workloads.
Large models and datasets can sometimes take a significant amount of time to download, especially during periods of high network utilization.
By default, each account can run only one instance in Non-GPU Mode at a time.
If you do not need GPU compute while downloading files, you can start the instance in Non-GPU Mode and complete the download first. Once the required models, datasets, or other files are ready, restart the instance with GPU resources and begin your workload. This helps avoid occupying and paying for GPU resources while simply waiting for large downloads to complete.
GPU availability is not reserved in Non-GPU Mode.
When you switch an instance to Non-GPU Mode, the GPU hardware previously used by your instance is released and may be allocated to other users. As a result, there is no guarantee that the same GPU, or even an available GPU, will still be available when you are ready to restart the instance with GPU resources. We recommend using Non-GPU Mode for long downloads or environment preparation only when you are comfortable releasing your current GPU allocation.