Training
Allocate memory for weights, optimizer state and activations. Multi-GPU training also needs a suitable interconnect.
HOPPER · DATA CENTER
NVIDIA H100 for demanding AI workloads
80 GB of HBM3 memory for workloads where the memory budget shapes the architecture. Plan your training or inference server around the model, dataset and GPU topology.
View configurationsYOUR SERVER
Build around your workload
GPU memory is one part of the server. Specify CPU, system RAM, storage and region to check the complete configuration.
- Model size and memory budget
- CPU, RAM and dataset storage
- Region and deployment requirements
WORKLOADS
When memory matters most
LLM inference
Plan VRAM for the model and KV cache together. Context length and concurrent requests influence memory demand.
Scientific computing
Match the numerical precision and CUDA software requirements to your calculation before choosing the host.
LOCATION
Choose a region for your workload
Keep compute close to your data and users. Capacity is checked for the selected model and location.
Europe
Availability on request
USA
Availability on request
Asia
Availability on request
COMPARE
Consider a different memory budget
Ampere80 GB
NVIDIA A100
Established CUDA workloads and model training
- Memory
- HBM2e
- Status
- On request
Price on requestConfiguration and region to be confirmed
Explore model Hopper141 GB
NVIDIA H200
Memory-intensive LLM inference and scientific computing
- Memory
- HBM3e
- Status
- On request
Price on requestConfiguration and region to be confirmed
Explore model Ada Lovelace24 GB
NVIDIA RTX 4090
Prototyping, image generation and rendering
- Memory
- GDDR6X
- Status
- On request
Price on requestConfiguration and region to be confirmed
Explore model Which H100 variant is described?
The reference specification is H100 SXM 80 GB. PCIe and NVL variants differ; the precise GPU variant must be confirmed in the server configuration.
Does 80 GB guarantee my model will fit?
No. Weights, precision, activations and cache share the memory budget. Check your framework and test the intended workload.
Can I request several GPUs?
Yes, include GPU count and interconnect requirements in your request. Capacity and topology need confirmation before deployment.
YOUR CONFIGURATION
Request an infrastructure quote
Choose resources and operating time. We will confirm a price for your configuration.
Each resource has its own listed configuration. Multiple GPU servers do not form an NVLink cluster. Monthly rates cover the full billing period; storage and traffic outside the configuration are not included.