AMD Instinct MI300X 192GB
USD 192.863 days · 1 lot · 1 GPUs · 192 GB per GPU
Configure 3 daysView full detailsIteraGPU / GPU configurations / AMD Instinct MI300X 192GB
Location GPU / AMD Instinct MI300X 192GB
MI300X: 192 GB for your AMD campaign.
Rent 192 GB on a single AMD card with the IteraGPU MI300X. A capacity to consider for your ML campaign when your recipe and dependencies are prepared for ROCm.
From USD 192.863 days · 1 lot · 1 GPUs · 192 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
AMD Instinct MI300X 192GB
1 GPUs included · 192 GB per card · HBM3
Package total · 3 days
USD 192.868 cards available.
Rent this configuration- GPUs per lot
- 1
- Memory per GPU
- 192 Go
- Memory type
- HBM3
- Available lots
- 8
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.
AMD Instinct MI300X 192GB
USD 450.007 days · 1 lot · 1 GPUs · 192 GB per GPU
Configure 7 daysView full detailsAMD Instinct MI300X 192GB
USD 1,590.0030 days · 1 lot · 1 GPUs · 192 GB per GPU
Configure 30 daysView full detailsPutting the 192 GB to work for a recipe prepared for AMD
The MI300X is a candidate when you're looking for more memory on a GPU, especially to keep a workload on a single card if its full peak fits within that envelope. For fine-tuning, count the weights, activations, and training states; for generative evaluation, add the cache tied to context and concurrent requests.
Prepare the software path precisely: PyTorch version for ROCm, custom extensions, attention, and the quantization method actually supported. A dependency that is exclusively CUDA may require a replacement or lead you to choose an NVIDIA card. Choosing an environment in the order expresses a preparation preference; it does not prove that your stack is already installed and validated.
The order can be paid directly in crypto, with no prior top-up. The process involves no KYC or ID document, with an account created from first name, last name, email, and a password.
- Recipe already validated on ROCm: compare its memory peak against the 192 GB per card.
- Project coming from CUDA: identify the components to adapt before choosing the rental period.
- AMD/NVIDIA comparison: keep the same data and quality criteria before interpreting a difference in duration.
Campaign example · 7 days
One week to evaluate an adaptation on a 192 GB card
The budget is USD 450.00 for a batch of MI300X, meaning one 192 GB GPU for 7 days. Prepare a minimal run with the chosen ROCm dependencies, then organize the campaign around output checks, fine-tuning, and its evaluation. Set aside time in this schedule for a possible port and exporting the results; the chosen period does not guarantee they will be completed.
USD 450.00
7 days · 1 lot · 1 GPUs · 192 GB per GPU
Configure this 7-day packageTwo alternatives to compare depending on your workload.
NVIDIA H200 SXM 141GB
USD 640.007 days · 1 lot · 1 GPUs · 141 GB per GPU
Worth considering if your project depends on CUDA and if its full peak fits within 141 GB per GPU. You are then comparing a lower per-card capacity against the ability to keep your NVIDIA software path.
Configure 7 daysView full details2 × NVIDIA B200 SXM
USD 2,071.007 days · 1 lot · 2 GPUs · 180 GB per GPU
Worth considering for a project you know how to split across multiple GPUs: a B200 batch contains two cards of 180 GB each. This is not a single 360 GB space; this option involves organizing the workload differently.
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.
Is the MI300X a good candidate for fine-tuning a model with ROCm?
The IteraGPU MI300X is worth considering for a fine-tuning recipe compatible with AMD/ROCm that needs a large memory envelope on one card. Concretely check the PyTorch, attention extensions, and quantization used by your project. The hardware offered does not guarantee that these software components are installed or already validated together.
How much does an MI300X rental cost at IteraGPU?
A batch of MI300X costs USD 192.86 for 3 days, USD 450.00 for 7 days, or USD 1,590.00 for 30 days. Each batch includes one 192 GB GPU. The amount covers the entire chosen period; a 7-day project on a batch therefore commits USD 450.00 of rental.
Does 192 GB guarantee that a model will stay on a single card?
The MI300X capacity is not enough to guarantee it. The peak depends on the weights and their precision, but also on activations, the optimizer, or the generation cache depending on the task. A workload can stay on one card if its full run fits within the envelope; verify this condition with your recipe.
When should you prefer an H200 SXM over the MI300X?
Compare with the H200 SXM if you want to keep a CUDA recipe, especially when an essential extension has no suitable AMD variant. The IteraGPU H200 SXM has 141 GB per GPU, versus 192 GB for the MI300X. This choice assumes your workload fits within 141 GB; the memory difference alone does not allow ranking their performance.
Your protocol / your measurements
Prepare a campaign you can compare.
Treat porting as an experiment
Inventory dependencies before renting: PyTorch version, custom operators, attention library, quantization and any inference engine. Check their support for ROCm and the MI300X, as well as the system compatibility matrix. The fact that a project uses PyTorch doesn't guarantee that all its CUDA extensions have an AMD variant. Start with a minimal batch that runs the model end to end.
Use the capacity without forgetting temporary states
The 192 GB can avoid model sharding when a smaller configuration forces you to split it. This simplification is worth testing before optimizing exchanges between multiple cards. First compute the weights, then measure the activations, the generation cache and the buffers actually used. For training, include gradients and the optimizer; for inference, vary context and concurrency.
A CUDA/ROCm comparison starts with parity: same data, same checkpoint, same generation parameters and a defined numerical tolerance. Examine the outputs and the business metric before comparing times.
When to prefer another configuration
If your method depends on a kernel that is exclusively CUDA and hard to replace, an H200 with 141 GB can reduce the adaptation work. Conversely, a project already validated on ROCm can take advantage of the MI300X's capacity without moving its software stack. Decide based on the total cost of the experiment, including validation hours and the number of repetitions, rather than on memory alone.
Plan the first AMD cycle
Reserve three days to validate installation, execution and outputs; seven days for a stable comparison with your baseline; thirty days for a campaign once the recipe is locked in. In the order, choose "Other environment" for your ROCm project and select the duration. Prepare a dependencies file and control data. Payment is made in crypto without an ID document request; your details are used for tracking the order.
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.
AMD Instinct MI300X 192GB · USD 192.86 for 3 days · 1 lot · 1 GPUs · 192 GB per GPU.