Tier 1
$49 / month
- 20 benchmark runs a month
- L4, RTX 4090 and RTX 5090
- Up to 2 GPUs per test, 1 test at a time
Plans from $49 a month
Know your AI model is ready for the GPU before you deploy it
Upload your model. We run it on real cloud GPUs and send you one clear report: does it work, are the answers right, and which GPU gives you the speed you need.
Add your model first, choose a plan when you start · No setup · Your model stays private
Deploying a model shouldn't be a gamble
“It worked on my laptop.”
Then it fails, slows down or behaves differently on the server GPU, after you have paid for it.
“Which GPU do I actually need?”
Rent too big and you overpay every month. Rent too small and you miss the speed your users expect.
“Is the faster setting safe?”
Speed-ups can quietly change your model's answers. You want to know before your customers do.
“Setting up GPUs takes days.”
Drivers, versions and runtimes stand between you and a single trustworthy number.
How it works
Upload your model
Send us the model file. No code, no setup, no explanation of how it works.
We test it on real GPUs
We run it on the GPUs you are considering, the same hardware you would rent, and check every answer.
Get your report
A clear verdict, the speed on each GPU, and what to change, in plain words.
What your report tells you
Does it work?
Whether your model runs on each GPU, and if not, exactly why.
Are the answers right?
Every GPU result checked against your model's original answers.
How fast is it?
Response time for one request, and how many it can handle per second.
Which GPU fits?
A side-by-side comparison of the GPUs you chose.
What to change
Simple advice when a setting is unsafe or a part of your model holds it back.
See a real report
Real results from our tests on NVIDIA H100, A100, RTX 5090, RTX 4090 and L4 GPUs (13–14 October 2026), using well-known public models, with estimates for H200 and B200. Your report looks just like this.
Sorts a photo into 1,000 categories.
Ready for GPU
Works on every GPU tested, and the answers match the original.
- Tested with
- one image of 224 × 224 pixels, one request at a time and in batches up to 64
- Answers
- Very close to the original at half precision; exact otherwise
- Fastest for one request
- RTX 5090 · 0.31 ms
- Most requests per second
- H100 · 76,799
GPU comparison
| GPU | Runs | Best setup | Response time | Requests per second |
|---|---|---|---|---|
| B200 Estimate180 GB | Not tested yet | GPU, optimizedTensorRT, half precision | 0.30 ms | 138,239 |
| H200 Estimate141 GB | Not tested yet | GPU, optimizedTensorRT, half precision | 0.34 ms | 88,319 |
| H100 Measured79.6 GB | Works | GPU, optimizedTensorRT, half precision | 0.36 ms | 76,799 |
| A100 Measured80 GB | Works | GPU, optimizedTensorRT, half precision | 0.72 ms | 41,731 |
| RTX 5090 Measured31.8 GB | Works | GPU, optimizedTensorRT, half precision | 0.31 ms | 58,890 |
| RTX 4090 Measured24 GB | Works | GPU, optimizedTensorRT, half precision | 0.64 ms | 45,099 |
| L4 Measured22.5 GB | Works | GPU, optimizedTensorRT, half precision | 0.53 ms | 10,890 |
Response time: one request on its own (median). Requests per second: with requests grouped in batches of up to 64. Estimates are scaled from the measured GPUs by published compute and memory speed, until we have tested those GPUs.
Test machines
| GPU | GPU memory | CPU | CPU cores | RAM |
|---|---|---|---|---|
| B200 | 180 GB | Not tested yet | ||
| H200 | 141 GB | Not tested yet | ||
| H100 80GB HBM3 | 79.6 GB | AMD EPYC 9554 64-Core | 61 | 242 GB |
| A100 SXM4 80GB | 80 GB | AMD EPYC 7V12 64-Core | 30 | 342 GB |
| RTX 5090 | 31.8 GB | AMD EPYC 9734 112-Core | 86 | 116 GB |
| RTX 4090 | 24 GB | AMD EPYC 7702P 64-Core | 30 | 126 GB |
| L4 | 22.5 GB | Intel Xeon Silver 4114 @ 2.20GHz | 14 | 47 GB |
CPU cores and RAM: what the test machine had available. The CPU-only results ran on that CPU.
What the report found
- On a GPU this model answers about 11× faster than on a CPU.
- GPU, standard on H100: answers very slightly different from the original. The report measures how much, so you can decide if that is acceptable.
- GPU, half precision on L4: slower than the standard GPU setup for this model, so do not use it. The fastest setup is 5× faster.
Every setup on each GPU
H100 · driver 535.309.01 (CUDA 12.4)
| Setup | Answers | Response time | Slowest 5% | Requests per second |
|---|---|---|---|---|
| CPU only | Same as original | 3.3 ms | 3.4 ms | 471batches of 16 |
| GPU, standard | Very close | 0.76 ms | 0.76 ms | 11,568batches of 32 |
| GPU, half precision | Same as original | 0.61 ms | 0.62 ms | 20,413batches of 64 |
| GPU, optimizedTensorRT, half precision | Same as original | 0.36 ms | 0.36 ms | 76,799batches of 64 |
A100 · driver 580.173.02 (CUDA 13.0)
| Setup | Answers | Response time | Slowest 5% | Requests per second |
|---|---|---|---|---|
| CPU only | Same as original | 3.6 ms | 3.7 ms | 331batches of 8 |
| GPU, standard | Same as original | 1.5 ms | 1.5 ms | 5,002batches of 64 |
| GPU, half precision | Same as original | 0.79 ms | 0.81 ms | 10,064batches of 64 |
| GPU, optimizedTensorRT, half precision | Same as original | 0.72 ms | 0.73 ms | 41,731batches of 64 |
RTX 5090 · driver 580.173.02 (CUDA 13.0)
| Setup | Answers | Response time | Slowest 5% | Requests per second |
|---|---|---|---|---|
| CPU only | Same as original | 4.3 ms | 4.4 ms | 352batches of 16 |
| GPU, standard | Same as original | 1.0 ms | 1.0 ms | 13,751batches of 64 |
| GPU, half precision | Same as original | 0.66 ms | 0.67 ms | 21,052batches of 64 |
| GPU, optimizedTensorRT, half precision | Same as original | 0.31 ms | 0.31 ms | 58,890batches of 64 |
RTX 4090 · driver 595.71.05 (CUDA 13.2)
| Setup | Answers | Response time | Slowest 5% | Requests per second |
|---|---|---|---|---|
| CPU only | Same as original | 3.4 ms | 3.4 ms | 352batches of 4 |
| GPU, standard | Same as original | 0.79 ms | 0.81 ms | 9,733batches of 64 |
| GPU, half precision | Same as original | 1.2 ms | 1.2 ms | 11,670batches of 64 |
| GPU, optimizedTensorRT, half precision | Same as original | 0.64 ms | 0.65 ms | 45,099batches of 64 |
L4 · driver 595.84 (CUDA 13.2)
| Setup | Answers | Response time | Slowest 5% | Requests per second |
|---|---|---|---|---|
| CPU only | Same as original | 8.5 ms | 9.3 ms | 167batches of 16 |
| GPU, standard | Same as original | 1.4 ms | 1.5 ms | 2,266batches of 64 |
| GPU, half precision | Same as original | 2.6 ms | 2.6 ms | 2,810batches of 64 |
| GPU, optimizedTensorRT, half precision | Same as original | 0.53 ms | 0.55 ms | 10,890batches of 64 |
About the model and the inputs we used
- Size
- 46.8 MB
- Parameters
- 11,699,112
- Format version
- ONNX opset 8
- Exported by
- —
| Input | Shape the model accepts | Shape we used | What we filled in |
|---|---|---|---|
data | N × 3 × 224 × 224 | 1 × 3 × 224 × 224 | normalized image values (about -2.5 to 2.5) |
Outputs: resnetv15_dense0_fwd (N × 1000). The full report is also available as JSON for engineers.
Built for teams putting AI into production
AI startups
Launch your AI feature knowing it will hold up, without a GPU expert on the team.
ML engineers
Get production numbers for a new model version in one step, instead of a week of setup.
Teams choosing hardware
Pick a GPU based on how your own model performs, not on spec sheets.
Why GridPup
- Real GPUs, not simulations. Every number comes from your model running on the actual hardware.
- Many GPUs in one place. GridPup already tracks cloud GPUs across providers, so you can test on the ones you would really rent.
- Results you can trust. Our speed measurements match NVIDIA's own benchmark tools on the same GPU.
- Private by default. Your model is used for your test only.
Plans
Pay monthly, cancel any time. A run is one model tested on one GPU, with the full report.
Tier 2
$249 / month
- 100 benchmark runs a month
- Every GPU up to H200 and B200 (bigger GPUs count as more runs: A100 and H100 3, H200 6, B200 10)
- Up to 5 GPUs per test, 3 tests at a time
Enterprise
Talk to us
- Runs to match your team
- Every GPU, including B300 and dedicated machines
- Large models, more tests at a time, API access
Questions
What models can I test?
We start with ONNX, the format PyTorch, TensorFlow, scikit-learn and most other tools can export to: image, text, speech, recommendation and tabular models. More formats will follow.
Can I benchmark generative AI models?
Yes, exported to ONNX: language models, speech models like Whisper, image and text encoders. We check that every part runs on the GPU, that the answers match the original, and how fast each step of the model is on each GPU. Tokens per second for complete text generation is coming next.
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, on request.
Do I need my own GPU or cloud account?
No. We provide the GPUs, set them up and run everything. You upload a model and get a report.
Is my model kept private?
Yes. Your model is used only for your benchmark and is not shared with anyone.
Do I have to tell you how my model works?
No. We work out what your model needs as input. If you have a few real examples, you can add them for an even more realistic test.
How do plans work?
Each plan gives a number of benchmark runs every month. A run is one model tested on one GPU. Runs reset each month, and you can cancel any time. If a test fails on our side, it does not use your runs.
What if I need more?
Tier 2 gives five times the runs, every GPU and three tests at a time. For larger teams, dedicated machines or API access, talk to us about Enterprise.
Talk to us
For Enterprise (more runs, dedicated machines, large models or API access) or any question about benchmarking your models, tell us what you need and we will get back to you.
Questions? Contact us.