GPU Selection - Compute Desk

GPU naming

The Compute Clear API uses two levels of GPU identification:

Level	Example	Scope
Family name	`H100`	Matches all variants of H100 (SXM5, PCIe, etc.)
Canonical name	`nvidia-h100-sxm5-80gb`	Matches one exact GPU variant

Use family names for broad searches. Use canonical names when you need a specific interconnect or VRAM configuration.

Discovering available GPUs

curl https://supply-api.compute-desk.com/available_gpus \
  -H "Authorization: Bearer $TOKEN"

The response maps family names to their canonical variants:

{
  "H100": ["nvidia-h100-sxm5-80gb", "nvidia-h100-pcie-80gb"],
  "A100": ["nvidia-a100-sxm4-80gb", "nvidia-a100-pcie-80gb", "nvidia-a100-pcie-40gb"],
  "L40S": ["nvidia-l40s-pcie-48gb"]
}

Using GPU types in offers

Both formats work in /get_offers:

# Family name — matches all H100 variants
compute:
  gpu_type: H100

# Canonical name — matches only SXM5 80GB
compute:
  gpu_type: nvidia-h100-sxm5-80gb

Popular GPU families

Family	Use case	Typical price range
H100	Large model training, inference	$2-5/GPU/hr
H200	Next-gen training, high-memory	$3-6/GPU/hr
A100	Training, fine-tuning	$1-3/GPU/hr
L40S	Inference, mixed workloads	$1-2/GPU/hr
L4	Cost-effective inference	$0.30-1/GPU/hr
RTX 4090	Budget training, inference	$0.30-0.80/GPU/hr

Prices vary by vendor, region, and contract type. Use /get_offers with max_price_per_gpu_hour to find options within your budget.

Filtering by vendor and region

Combine GPU selection with vendor and region filters:

vendor: nebius
region: EU
compute:
  gpu_type: H100
  gpu_count: 8
  max_price_per_gpu_hour: 4
  contract_type: ondemand

Offer Lifecycle Cloud-Init

​GPU naming

​Discovering available GPUs

​Using GPU types in offers

​Popular GPU families

​Filtering by vendor and region

GPU naming

Discovering available GPUs

Using GPU types in offers

Popular GPU families

Filtering by vendor and region