Find out which GPU your model needs before you pay for one. GridPup runs your model on real cloud GPUs and reports whether it works, whether the answers are right, and how fast it is on each GPU. You need no GPU or cloud account of your own.
Sample result: Qwen2.5-7B on four GPUs
Qwen/Qwen2.5-7B-Instruct served with vLLM, 512 tokens in and 256 out per request, with 1 to 128 users at once. “Users served” counts users whose first token arrives within one second for 95% of requests.
| GPU | Tokens/s, one user | Users served |
|---|---|---|
| H100 80 GB | 107.9 | 32 |
| A100 80 GB | 96.9 | 32 |
| RTX 5090 32 GB | 103.7 | 8 |
| L40S 48 GB | 49.8 | 8 |
Measured by GridPup in October 2026, one copy of the model per machine, vLLM default settings.
What the report tells you
Does it work?
Whether your model loads and runs on each GPU, and whether every part of it runs on the GPU rather than falling back to the CPU.
Are the answers right?
Whether the answers on the GPU match your original model, including with faster settings such as half precision.
How fast is it?
Tokens per second for one user and under load, time to the first token, and how many users each GPU serves within your target.
Which GPU fits?
The GPUs you chose side by side, so you pick the one that meets your needs instead of the biggest one.
How it works
- Give your model. Enter a Hugging Face chat model by name, or upload an ONNX file. GridPup works out the inputs your model needs; you can add a few real examples.
- Pick GPUs. L4, RTX 4090 and RTX 5090 on every plan; A100, H100, H200 and B200 on Tier 2.
- GridPup tests it. It sets up each GPU, serves your model and sends it real requests, from one user to many at once.
- Read the report. One page per model with the results on every GPU and what to change. Your model is used only for your benchmark.
Plans
Tier 1: $49 a month
20 benchmark runs a month on L4, RTX 4090 and RTX 5090. Up to 2 GPUs per test, one test at a time.
Tier 2: $249 a month
100 runs a month on every GPU up to H200 and B200 (bigger GPUs use more runs). Up to 5 GPUs per test, 3 tests at a time.
Enterprise
B300 and other GPUs, dedicated machines and API access. Talk to GridPup.
A run is one model tested on one GPU. Runs reset every month; cancel any time.
Already serving your model?
Test the endpoint you run today from the US, Europe and India: time to first token, tokens per second and behaviour under load. See LLM APIs and hosting.
Frequently asked questions
How many tokens per second does an H100 give for a 7B model?
In GridPup's test of Qwen2.5-7B-Instruct on vLLM (512 tokens in, 256 out), one H100 gave 108 tokens per second to a single user, served 32 users at once with the first token within a second, and reached about 5,800 tokens per second in all with 128 users.
Is an RTX 5090 good for LLM inference?
For one user it is almost as fast as an H100 on a 7B model (104 against 108 tokens per second in the same test). With its 32 GB of memory it serves fewer users at once: 8 users within a 1 second first token against 32 on an H100 or A100.
What models can I benchmark?
Chat models from Hugging Face by name, served with vLLM and timed in tokens per second, and ONNX files exported from PyTorch, TensorFlow or scikit-learn: image, text, speech, recommendation and tabular models. Licence-gated models are not supported yet.
Which GPUs can I test on?
L4, RTX 4090 and RTX 5090 on every plan; A100, H100, H200 and B200 on Tier 2. B300 and other GPUs are available on Enterprise plans.
How much does a benchmark cost?
Plans start at $49 a month for 20 benchmark runs; Tier 2 is $249 a month for 100 runs and every GPU up to H200 and B200. A run is one model tested on one GPU. If a test fails on GridPup's side it does not use your runs.