GPU servers
Comparing GPU servers without a pretend benchmark.
GPU models, allocated VRAM, access, hourly billing and an honest memory-capacity value metric. How to compare GPU instances and bare-metal hosts.
By RackRadar · · 3 min read
GPU servers need a different comparison from CPU-only hosting. An inexpensive host with a large disk is not automatically a good place to run a GPU workload. Start with the accelerator your software supports, the memory your workload needs, and the kind of access you receive.
GPU capability and host allocation
A GPU can belong to a virtual machine or a bare-metal server. RackRadar keeps those as separate properties. Host allocation describes the server; GPU access distinguishes a full device, a partition, a shared GPU, or an unknown allocation. A GPU count alone cannot tell you whether devices are shared.
We keep exact published model names. Two cards with the same VRAM can have very different throughput, memory bandwidth, precision support, power limits and software compatibility. We do not manufacture benchmark scores from their names, CUDA core counts or advertised TOPS.
What VRAM value means
The GPU view uses a separate capacity metric:
VRAM value = total allocated VRAM in GB × 100 ÷ recurring monthly USD.
For a hypothetical 48 GB allocation costing $600 per month, the result is 8 GB per $100/month. At $1,200, the same allocation scores 4. Higher means more published memory capacity per dollar. It does not mean faster training, inference or rendering. This is version 1 of the GPU capacity methodology.
Compare the same GPU model and access type before using this metric. We show whole-host cost and monthly price per advertised GPU alongside it. Unknown memory or GPU count stays unknown and cannot produce that corresponding metric. GPU offers do not receive the CPU advertised-value index.
Total memory is not one memory pool
A four-GPU plan with 96 GB per card has 384 GB total VRAM. That does not mean one process can use a single 384 GB device. Your framework, model parallelism and the interconnect determine how you can use multiple GPUs. RackRadar does not infer NVLink, a fast fabric or automatic memory pooling.
Read the original billing terms
The primary comparison amount is recurring monthly USD. It normalizes the published monthly or period price, then converts that recurring amount using RackRadar’s fixed reference exchange rates. Setup fees remain separate one-time charges and do not enter monthly price or capacity value. Unknown fees are not proven free.
Hourly prices remain in their original currency, including fractional cents. When a provider supplies a monthly reference amount, we label it as a reference, not a guaranteed cap. We do not silently multiply an hourly price by 730 or promise a monthly maximum. Required extras, tax treatment, billing increments, idle charges, storage and transfer can affect your actual bill. Check the provider’s terms; missing billing details remain unknown.
Where the first GPU offers come from
The initial collection covers Hetzner’s published default GEX configurations and Vultr plans with explicit structured GPU model, count and allocated-memory fields. Hetzner’s descriptions supply GPU memory separately from system RAM; optional upgrades are not counted as included capacity. Vultr’s VRAM field is the plan’s total allocation. We do not decode missing hardware facts from opaque plan IDs.
A catalog location is not live stock. A missing location list means unpublished locations, not that the server is available everywhere. Wider provider and variant coverage will follow source evidence rather than guessed specifications.