L40S
coming soon.
48 GB with Ada tensor cores and FP8. A balanced card for inference and image generation that does not need data-centre memory bandwidth.
Not available to rent yet
What the L40S
is built for.
Tensor-core precisions: FP8, BF16, FP16, TF32.
Best at
- Flux and SDXL image generation
- Serving 8B–13B models in FP16
- Video and rendering workloads
- FP8 inference
Coming soon
The L40S is not on sale yet. Tell us what you would run on it and we will let you know when it launches.
Get notifiedWill my model fit
on 48 GB?
Rules of thumb for weights plus a modest KV cache or optimizer state. Longer contexts and bigger batches need more.
| Workload | Needs | On a L40S |
|---|---|---|
| Llama 3.1 8B · FP16 inference | ~18 GB | Fits |
| Qwen 32B · FP8 inference | ~36 GB | Fits |
| Llama 70B · 4-bit inference | ~42 GB | Tight |
| Llama 70B · FP8 inference | ~78 GB | Too big |
| Llama 8B · LoRA fine-tune (BF16) | ~24 GB | Fits |
| 32B · QLoRA fine-tune | ~40 GB | Fits |
| Flux.1 [dev] · BF16 | ~33 GB | Fits |
| SDXL · FP16 | ~12 GB | Fits |
Guides for
the L40S.
Benchmarks and sizing guides from our blog.
L40S GPU Cloud: Choosing the Right GPU for Production Inference
Most GPU comparisons start with the biggest card and work backward to justify it. L40S starts with what inference actually needs, and prices infrastructure accordingly.
Not the right fit?
Compare the fleet.
These are on sale now, billed hourly with no contract.
The L40S
is coming soon.
Tell us what you would run on it and we'll let you know when it launches.
Get notified
