Compare a setup, not just a card.

Set the criteria before looking at the price

A useful comparison starts with a short common record. Fix the application version, its backend, the input set, the precision, the work batch sizes and the quality criterion. If you change any of these parameters, record it as a separate experiment. Then keep the reason each candidate was eliminated or selected.

For a migration to AMD, consult the ROCm matrix matching your hardware and your software stack. Published compatibility is a prerequisite; it does not prove that your application or a requested environment preparation already works. Acceptance must verify the versions and the result on the target.

Distinguish lots, cards and memory per GPU

A lot is the rental unit. The number of cards depends on the configuration: a B200 lot contains two cards, while an RTX 4090 lot contains one card. The number of cards requested is calculated as follows: number of lots × cards per lot. The total in USD is calculated using the lot price and the chosen duration, without multiplying a second time by the number of cards.

The memory indicated is that of each GPU. Two 180 GB B200s represent two 180 GB memory spaces; their arithmetic sum is not memory automatically available to a task. NVIDIA describes the software and hardware conditions for multi-GPU exchanges. The topology and links of the machine concerned must be observed, not inferred from the model name.

Keep the units of your readings. A decimal GB corresponds to one billion bytes; a GiB corresponds to 2 to the power of 30 bytes. A discrepancy between the spec sheet and a tool displaying binary units must be explained before concluding that memory is lacking. The advertised capacity also does not give the free memory during your run.

Compare a cost under identical financial conditions

Here is an example budget, taken from our catalog on September 24, 2026. A team is looking at a single lot for seven days, with a single card. It provisionally assumes a workload that fits within 24 GB, subject to its test. This assumption opens up two options to examine; it gives them neither the same speed, nor the same software support, nor the same result.

For this scope, the price difference is 110.00 − 55.00 = 55.00 USD. This is a difference in commitment over seven days. Without measured time for an accepted result, the cost per completed task cannot be inferred. The team's preparation time and any external transfer or payment fees must be examined separately; they are not costed here.

Keep a decision the team can accept

End the comparison with four lines: option chosen, determining criterion, accepted limit and remaining check to perform. If results are missing, mark the decision as provisional. A larger capacity is only relevant if it solves an identified constraint; a lower price does not prove that the task will cost less to complete.

Acceptance file v1 includes a grid of ten checks, a decision template, documented inventory commands and a fictional example. Fill in your inputs, observations with units and the evidence supporting each status. These documents contain no GPU measurements. They help organize the evidence your team will need to gather.

Shortlist candidates, then clear the conditions

To recommend a configuration to your team, start with a short list of eligible candidates. A mandatory criterion that is not met rules out an option; a mandatory criterion that is unknown leaves it to be confirmed. Price, familiarity with a tool or headroom then separate the options that cover the need. A preference does not make up for a missing requirement.

The same sheet is used to look for an alternative to your current environment: keep the constraints that must survive the change and name those you are willing to modify. Asking for more cards does not by itself solve a problem of queueing, data preparation or available duration.

Two situations for building your selection

First situation, hypothetical: your independent task requires 40 GB per GPU, margin included, based on your own measurement and the same units as the spec sheet. Nominal capacities below that are ruled out. The 48 GB L40S, RTX 6000 Ada, RTX A6000 and A40 may join the list to examine: each lot contains one card. This list covers the four 48 GB options in the catalog; it does not rank their speed and does not exclude higher capacities. Software, actually free memory and machine resources still need to be verified.

Second situation: an application must distribute the same processing across two GPUs. The B200 lot provides two cards to examine, but its name alone is not enough to validate this scenario. Document the distribution, the versions and the required exchanges, then compare them with the proposed machine. If your project requires a specific topology that is still unknown, the selection remains provisional. Two independent tasks constitute another scenario, with a memory check and an accepted result for each.

Know when to suspend the decision

This option is less suitable for a team expecting a fully managed application, immediate preparation or an already guaranteed availability commitment. Choosing an environment in the order does not prove a working installation. If these points are essential, record the information or commitment to be obtained before selecting the offer.

Acceptance then verifies the agreed target and test case. A successful launch validates neither all datasets nor future periods. In the grid, distinguish compliant, deviation and unverified; a mandatory check not performed prevents declaring the scope accepted. Keep the evidence, the accepted limit and the decision-maker's name, with an explicit reservation or a refusal depending on the deviation.