NVIDIA H100 SXM 80GB
USD 203.573 days · 1 lot · 1 GPUs · 80 GB per GPU
Configure 3 daysView full detailsIteraGPU / GPU configurations / NVIDIA H100 SXM 80GB
Location GPU / NVIDIA H100 SXM 80GB
H100 SXM 80 GB: rent the capacity for your Hopper campaign.
Rent an 80 GB H100 SXM from IteraGPU for your Hopper-ready fine-tuning or training campaign. Choose the length of your comparisons, with a complete budget per lot.
From USD 203.573 days · 1 lot · 1 GPUs · 80 GB per GPU
Crypto without KYC. Account required: first name, last name, email and password. No KYC procedure and no ID document.
Your rental3 / 7 / 30 days
NVIDIA H100 SXM 80GB
1 GPUs included · 80 GB per card · HBM3
Package total · 3 days
USD 203.5763 cards available.
Rent this configuration- GPUs per lot
- 1
- Memory per GPU
- 80 Go
- Memory type
- HBM3
- Available lots
- 63
Choose the period for your campaign.
Price in USD for a full lot; from 1 to 10 lots in the configurator. The total is the plan multiplied by the lots, with no automatic renewal.
NVIDIA H100 SXM 80GB
USD 475.007 days · 1 lot · 1 GPUs · 80 GB per GPU
Configure 7 daysView full detailsNVIDIA H100 SXM 80GB
USD 1,690.0030 days · 1 lot · 1 GPUs · 80 GB per GPU
Configure 30 daysView full detailsChoosing Hopper with a defined memory requirement
Choose the H100 SXM when your campaign targets a Hopper-compatible software stack and 80 GB per card matches your memory budget. For fine-tuning, that budget includes the base model and the activations; for full training, add the gradients and optimizer states. The requirement for the same architecture can therefore change depending on the method you use.
The 80 GB can leave room for a microbatch or sequences that were constrained on a smaller configuration, but your recipe must establish that margin. If it already runs on Ampere and fits within 80 GB, the A100 SXM deserves a plan comparison. If capacity remains insufficient, look at the 141 GB H200 SXM rather than assuming that a different precision will solve the problem without validation.
- Define what will be trained: adapters alone or the full set of parameters, along with their associated states.
- Check the CUDA/PyTorch versions and required extensions before choosing this Hopper configuration.
- Compare the budget for the period with the expected results of the campaign, without inferring a runtime from the H100 name.
Campaign example · 7 days
Seven days to compare your training settings
The rental budget is USD 475.00 for 1 H100 SXM lot, i.e. 1 80 GB GPU for 7 days. Your program can bring together a baseline recipe, a few compatible microbatch or precision variants, then metric comparison and a checkpoint resume. Set your priority runs before booking: the plan defines a period, without guaranteeing how many variants your model will complete.
USD 475.00
7 days · 1 lot · 1 GPUs · 80 GB per GPU
Configure this 7-day packageTwo alternatives to compare depending on your workload.
NVIDIA A100 SXM 80GB
USD 280.007 days · 1 lot · 1 GPUs · 80 GB per GPU
Compare another 80 GB card when your recipe is already Ampere-compatible. The plan and the outcome of your protocol drive the decision; identical capacity does not guarantee the same behavior or the same throughput.
Configure 7 daysView full detailsNVIDIA H200 SXM 141GB
USD 640.007 days · 1 lot · 1 GPUs · 141 GB per GPU
Move on to studying a 141 GB card if memory, rather than the choice of precision, is preventing your experiment from fitting within 80 GB. The H200 still needs to be validated with your software and the budget for the period you choose.
Configure 7 daysView full detailsPay for your rental directly in crypto.
Direct payment with no prior top-up, including BTC on Bitcoin or USDT on Tron. Account required: first name, last name, email and password. No KYC procedure and no ID document.
View assets, networks, and payment termsChoose your rental: the useful answers.
Why choose the H100 SXM for a fine-tuning campaign?
The IteraGPU H100 SXM provides an 80 GB envelope on one GPU for a campaign whose recipe is Hopper-compatible. It is a candidate when the weights, activations, and states of the trained parameters require more room than a smaller card. You need to size your fine-tuning method; the GPU name guarantees neither that a given model will fit nor that fine-tuning will improve your metric.
What is the price of an H100 SXM lot for my research period?
IteraGPU offers 1 H100 SXM lot of 1 GPU with 80 GB for USD 203.57 over 3 days, USD 475.00 over 7 days, or USD 1,690.00 over 30 days. These are the lot prices for full periods. Choose a duration that includes your runs, their evaluation, and the export of results. Direct crypto payment does not require pre-funding a balance.
Is 80 GB on the H100 SXM enough as soon as the model weights fit?
No. On the H100 SXM, as on any other card, the size of the weights is only part of the required memory. Training also uses activations, gradients, optimizer states, and temporary allocations. Parameter fine-tuning reduces some of these items without eliminating the others. Check the peak of a complete step and that of evaluation before choosing the 80 GB lot.
Should you compare the H100 SXM with the A100 SXM or the H200 SXM?
Compare the A100 SXM when your recipe fits within 80 GB and runs on Ampere: the decision then comes down to plans and your protocol. Compare the H200 SXM when you are looking for more memory on a card, with 141 GB instead of 80 GB. At IteraGPU, each lot of these three models contains 1 GPU. None of these characteristics is enough to name a winner for every campaign.
Your protocol / your measurements
Prepare a campaign you can compare.
Evaluate precision alongside convergence
Hopper supports FP8 operations, but this format requires a matching software stack and model. Start with a BF16 or FP16 run validated by your project, then look at a reduced-precision variant. Keep the same evaluation examples and monitor loss, non-finite values and the final metric. A higher number of steps per second is only valuable if the experiment still answers its question.
Making the 80 GB work without skewing the protocol
Gradually increase the microbatch until you reach a reasonable memory margin, measuring a step that includes backpropagation. Gradient accumulation then lets you build a larger effective batch without requiring all its examples to be in memory at once. Log both sizes: they do not mean the same thing for performance or for the parameter update.
Before a long campaign, actually reload a checkpoint and run a few steps. This test reveals missing optimizer or learning-rate scheduler state earlier than a simple check that the file exists.
SXM, PCIe or more memory?
Keep a CUDA/PyTorch build compatible with Hopper and check your attention extensions. The H100 PCIe is a relevant comparator for single-card work, but a different hardware variant must be measured on your workload. Prefer looking at the H200 if memory capacity remains the main problem. The A100 SXM is also useful for comparing the cost of an experiment whose kernels and precision already work on Ampere.
Pairing the rental with a research deliverable
Over three days, aim for a batch ramp-up curve and a reusable checkpoint. Over seven days, compare a few well-chosen configurations. Over thirty days, organize repetitions and intermediate evaluations. Order with the duration, number of cards and environment you want; pay in crypto then click "I've paid". Your notebook can keep the exact recipe for each trial alongside the order tracking.
Prepare your first trial on this configuration
KV cache: calculating memory by context, GQA and batchComparing quantizations: memory, quality, and useful costMicrobatch and accumulation: computing the effective batch
The GPU spec describes a capacity. These methods help you define the inputs, the settings and the result to verify on your own workload.
Your campaign, your duration.
NVIDIA H100 SXM 80GB · USD 203.57 for 3 days · 1 lot · 1 GPUs · 80 GB per GPU.