Current GPU family
- NVIDIA RTX 5090 — 32 GB of GDDR7 VRAM per GPU
The Compute console is the source of truth for the presets you can launch at that moment. It shows the GPU count, vCPU, system RAM, disk, bandwidth, location, availability, and active price before you create an instance.
Multi-GPU instances
Compute can offer RTX 5090 configurations with multiple GPUs, up to 8 GPUs per instance when capacity is available. Choose the GPU count based on the workload rather than assuming that more GPUs always make a single job faster.- 1 GPU: A good starting point for inference, image generation, development, and workloads that fit within 32 GB of VRAM.
- Multiple GPUs: Useful for workloads that explicitly support tensor parallelism, distributed training, multi-GPU rendering, or several jobs running in parallel.
VRAM is attached to each GPU. Applications do not automatically treat multiple GPUs as one larger shared memory pool. Your framework or model must support multi-GPU execution to use them together effectively.
VRAM and system RAM
VRAM is memory on the GPU. It determines whether model weights, activations, image-generation pipelines, and other GPU data fit on a card. System RAM is separate memory used by the operating system, runtime, data loaders, preprocessing steps, and CPU-side parts of your workload. For example, one RTX 5090 provides 32 GB of VRAM. A four-GPU instance provides four GPUs with 32 GB each, but whether a workload can use all four depends on the software and how it distributes work.Choose a current preset
When you create an instance:- Select a location.
- Choose GPU under the setup options.
- Compare the presets currently available.
- Check the GPU count, VRAM, vCPU, RAM, disk, bandwidth, and price shown in the console.
- Choose the smallest configuration that comfortably fits the workload.