Compute and GPU profiles
Choose memory, GPU count, CPU, and system RAM for a service.
Supported GPU definitions
| API model | Profile memory per device | Allocation |
|---|---|---|
rtx-pro-6000-blackwell | 24 GB / 48 GB | One MIG device per service |
rtx-pro-6000-blackwell | 96 GB | Full GPU |
h200 | 141 GB | Full GPU |
b200 | 180 GB | Full GPU |
b300 | 288 GB | Full GPU |
These are supported product profiles, not a claim of live inventory or measured usable VRAM. Full-device profiles accept 1–8 GPUs, subject to capacity. MIG profiles accept one device. The model and memory must match exactly.
CPU and system memory
The shared resource contract accepts 1–64 vCPU and 1–1,024 GB system memory. Set gpuCount to 0 for CPU-only compute. Resource schema limits are not a promise that every placement can satisfy every size.
{
"gpuCount": 1,
"gpuModel": "h200",
"gpuMemoryGb": 141,
"cpu": 8,
"memoryGb": 32
}Availability and pricing
The placement must be enabled and connected, declare the selected GPU model, have sufficient capacity, and have a configured price for the selected profile. Missing prices make a GPU unavailable. CPU-only requests do not require a GPU allocation.
Resize a service
Use the console resize action or the API resize action with a complete resource object. Resizing can restart processes. The service endpoint and persistent workspace remain associated with the service. The old allocation must be released before a replacement is confirmed.
A resize request is asynchronous. Inspect state until it reaches the intended result; do not treat an acknowledgement as a running GPU.