Service facts
IteraGPU is the name of the service presented on this site. The points below describe our offering and the available tools. Detailed pricing and the stock of each configuration can be found in the catalog; a card's capacity alone does not describe the server or the result of your software.
| Reference point | Value or how it works | How to read it |
|---|---|---|
| Audience | ML research and engineering | Training, inference and fine-tuning according to your protocol |
| Catalog | 15 NVIDIA and AMD configurations | Choose based on compatible software, memory and measured load |
| Memory | 24 to 192 GB nominal per GPU | GB here refer to manufacturer GB; these are neither free GiB nor server RAM |
| Composition | 1 GPU per lot; 2 GPUs for B200 | The lot price includes its cards; their memory is not pooled automatically |
| Location | 3, 7 or 30 days; 1 to 10 lots | Full amount in USD per lot, multiplied by the quantity |
| Availability | Stock shown on each listing | Payment confirmation and access preparation are two separate steps |
| Software setup | Desired environment at checkout | A preference is not proof that your model is installed or compatible |
| Tracking | Account, orders and ML notebook | Notes you enter do not replace your files, backups or external measurements |
Technical sources: Catalog and plan specifications · Flat rates · Ordering terms
The GPU is a protocol decision
Memory must accommodate the planned work, the environment must suit the model, and the schedule must include setup, measurements, restarts and export. For inference, set the input and output lengths as well as the number of concurrent requests. For training, examine the full step and the states needed for resume.
The sizer computes an envelope from the parameters, the weight format and your execution reserve. That reserve is an explicit assumption; it is not a guaranteed capacity. To decide between two configurations, keep a common set of inputs and a common quality rule, then measure the cases that matter for your project.
Technical sources: Sizer and assumptions · Prepare inference requests
Original resources, with an explicit scope
IteraGPU Lab v1 brings together a compute and memory measurement notebook, two Python scripts, a quality protocol, plan pricing and a blank raw table. The original code and documentation are distributed under the MIT license. The instructions state the prerequisites and the explicit errors when PyTorch or the GPU is missing.
The measurement script was tested on an off-catalog RTX 5070 with an 8,320-parameter network. This test verifies the instrumentation; it measures neither a large model nor the performance of rented plans. The memory folder specifies the conditions and the results.
The memory method explains the counters; the benchmarking method compares results at fixed quality and the cost of the plan incurred. The distributed resources do not contain pre-filled notebook outputs or GPU results. Measure your workload before drawing a recommendation from it.
- Download IteraGPU Lab v1
Versioned archive: protocol, notebook, scripts, tables, instructions, inventory and license.
- Read the prerequisites and the validation scope
Prerequisites, tested parameters and limits of the test.
Technical sources: Estimation and peak memory · Comparison and cost per validated corpus
Autonomy in your workloads
You choose your software, your models and how you organize your compute. IteraGPU does not inspect the content of your files, prompts or workloads. This policy applies to the content of the work; it does not remove the information needed for the account, the order and payment tracking.
Prepare the dependencies and versions, check the usage rights and organize your backups. The environment selected at checkout expresses your preparation needs. Its compatibility with your libraries, extensions and models must be checked before launching the full campaign.
Technical sources: Data and privacy
A logbook to keep the reason for your choice
The ML logbook keeps the objectives, parameters and observations you enter. Link your hypothesis to a code version, a set of inputs and an order reference. Also keep the rejected variants: memory overflow, insufficient quality or incomplete results explain the next decision.
The logbook does not run your code, does not automatically collect GPU telemetry and does not upload your datasets. Keep the raw files, outputs and checkpoints in your own backups. A note should let you find the evidence, without replacing it.
From plan to rental tracking
Choose the model, the duration and the number of lots, then check the USD total. The price corresponds to the entire plan; it is not an hourly meter. Crypto payment is made without KYC or identity document request. The first name, last name and email are used to track the order.
After a transfer, "I have paid" records your report for verification. Confirmed payment and delivery of GPU access are tracked separately. Your account gathers orders, documents and support requests; check the readiness status before starting your workloads.
Technical sources: Crypto payment journey
Practical questions
Does IteraGPU provide a universal GPU recommendation?
IteraGPU provides criteria and tools to prepare your choice, without naming a winning GPU for all workloads. Memory, environment, context, concurrency and quality must be checked together. The budget is then compared between the options that meet these requirements.
Do two B200 GPUs form a single memory?
A B200 lot includes two GPUs, each with its own memory. Your software must explicitly distribute the data or compute to use them together. The composition of the lot does not demonstrate unified memory, a particular topology or a given acceleration.
Does choosing PyTorch at checkout guarantee that my model will work?
Choosing an environment expresses a preparation preference, not a validation of your model. Check the driver, libraries, extensions and required versions, then run a small representative case before your campaign.