Room for a heavier experiment
Plan resources for legitimate training, fine-tuning and repeatable model work.
Content revised · English
- Architecture to evaluate
- LoRA / QLoRA / full fine-tuning
- Prepared starting points
- 3 plans
- Responsibilities
- Application, rights and recovery
A fit shortlist, not installed software or live inventory.
Keep these in view.
- LoRA / QLoRA / full fine-tuning.
- Make the stack and data reproducible.
- Custom hardware is a request, not stock.
LoRA / QLoRA / full fine-tuning
Inference, adapter tuning and full fine-tuning have different memory demands. Weights are only part of training memory: activations, optimiser state and gradients also matter. Check the batch, sequence length, precision and checkpoint strategy on the actual experiment.
Make the stack and data reproducible
A project may use CUDA, PyTorch, Transformers, Accelerate or related tooling after checking compatibility. Record environment versions, dataset rights, random seeds and evaluation results. Stage data securely and keep checkpoints encrypted in a separate recovery location.
Custom hardware is a request, not stock
The prepared catalogue shows single-device configurations. A multi-GPU topology, NVLink arrangement, short burst term or custom delivery date needs an explicit proposal and compatibility review. There is no hourly scheduler or multi-H100 inventory connected.
Start with the experiment
Dataset size, model parameters, batch size and optimiser state determine the memory requirement. Check framework compatibility and the installed GPU driver before committing to a long run.
Keep work reproducible
Version code, environment definitions and dataset provenance. Write checkpoints outside temporary working directories, and record enough configuration to continue a failed job or reproduce a result.
Respect the data boundary
Training material can contain protected or personal information. Confirm rights, minimise sensitive input and control access to intermediate artefacts, logs and model outputs.
Capacity to consider.
Match these starting points to the representative application. Software installation and live inventory are not connected; the exact specification and period total come from the catalogue.
Spectre A6000
Use one NVIDIA RTX A6000 with 48 GB of GPU memory alongside 128 GB of ECC host memory and 2 TB of NVMe storage. The larger GPU memory allocation and 10 Gbps uplink suit more demanding inference or selected training workflows, with model and optimiser memory assessed before deployment.
- Compute
- 16 vCPU (EPYC)
- Memory
- 128 GB DDR5
- Storage
- 2 TB NVMe SSD
- Transfer
- 30 TB / month
billed monthly
Spectre H100
Build intensive accelerator workloads around one NVIDIA H100 with 80 GB of HBM3, 256 GB of host memory and 4 TB of NVMe storage. A 25 Gbps uplink and 50 TB monthly transfer allowance support large datasets; high-parameter models still require careful precision, context and memory planning.
- Compute
- 32 vCPU (EPYC)
- Memory
- 256 GB DDR5
- Storage
- 4 TB NVMe SSD
- Transfer
- 50 TB / month
billed monthly
Eclipse Dual
Reserve two Xeon Gold 6338 processors for a total of 64 physical cores and 128 threads, supported by 256 GB of ECC memory. Four 4 TB NVMe drives form RAID-10, with a 10 Gbps port, unmetered transfer and a 400 Gbps network mitigation tier for large parallel workloads.
- Compute
- 2× Intel Xeon Gold 6338 (32c / 64t each)
- Memory
- 256 GB DDR4 ECC
- Storage
- 4× 4 TB NVMe (RAID-10)
- Transfer
- Unmetered
billed monthly
Before this workload goes live.
- Check the exact plan’s CPU, memory, disk, transfer and port.
- Choose a listed region and verify the workload’s permitted use.
- Test the representative application and an independent restore.
Questions, answered.
Can I reserve several H100 devices?
A multi-device arrangement is a custom request to review, not listed live stock. Confirm interconnect, topology, framework compatibility and delivery before committing.
Is GPU billing hourly?
The prepared catalogue uses monthly, annual and biennial totals. No hourly billing scheduler is connected. A special burst term would need its own accepted proposal.
Can I train on personal information?
Dataset rights, a lawful processing basis, access, retention and transfers need assessment. A particular region or device does not create that permission.
Are frameworks and restricted-device permissions included?
Framework compatibility and any applicable export or access restriction must be checked for the actual operator and workload. No blanket exemption or preinstalled training environment is claimed.
What is kept in my account workspace?
Your configuration, selected plan, image, location and billing term are kept with the order and payment request references. Quotes, payment-state reads and support conversations remain linked to your account. Read fulfilment details separately from payment status.