ARVICA / DOCUMENTATION

Docs / Compute

GPU and instance selection

Select memory and topology for the workload you will run.

Memory before model name

Allow room for model weights, activations, runtime caches and batch size. Increasing the GPU count does not automatically create one shared memory pool. Check that your framework supports the intended parallel execution.

GPU, host and rack

Compare both the accelerator and host configuration. CPU architecture, RAM, local disk, PCIe or SXM packaging and interconnect matter. GB200/GB300 rack-scale configurations require a defined system scope rather than a single-card comparison.

VM and bare metal

Confirm whether the offered instance is a VM or bare metal and what administrator access is included. Do not infer root access, NVLink, InfiniBand or a particular host specification from a GPU family name alone.

Compare configurations