Once a job outgrows 24 GB consumer cards but does not need an 80 GB flagship, three cards come up again and again: the L40S, the A100 and the RTX 6000 Ada. The L40S vs A100 choice in particular confuses people, because the older A100 is famous for training while the L40S is newer and, on some spec lines, much stronger. This article compares the three for inference, fine-tuning and rendering, using only the figures shown on Cloudj's model pages, and explains how to search when your first choice is not listed.
For the broader trade-off between card classes, start with consumer vs datacenter GPUs in the cloud.
The three cards side by side
| L40S | A100 PCIe | RTX 6000 Ada | |
|---|---|---|---|
| Architecture | Ada Lovelace | Ampere | Ada Lovelace |
| VRAM | 48 GB class | 40 GB | 48 GB |
| FP32 (model page) | 73.3 TFLOPS | 15.6 TFLOPS | 81.4 TFLOPS |
| Memory bandwidth (model page) | 626 GB/s | 1313 GB/s | 764 GB/s |
| Power | 350 W | 250 W | 300 W |
| Class | Datacenter | Datacenter | Workstation |
What stands out: the two Ada cards have more VRAM and far higher FP32 throughput, while the A100 has roughly twice their memory bandwidth. That one difference explains most of the practical results below. Workloads that stream large weights through memory over and over (large-batch training, some LLM decoding) lean on bandwidth; workloads with heavy arithmetic per byte (rendering, image generation, many vision models) lean on compute.
The model pages also list an A100 SXM4 variant, the same 40 GB size with a higher power budget. All three classes need a higher account level than new accounts start with.
Inference
LLM serving. Token generation for a single stream is often limited by memory bandwidth, which favours the A100. But the A100's 40 GB holds less: a 30B to 34B model in 4-bit with a generous KV cache is comfortable on 48 GB and tight on 40 GB. For many teams, the extra 8 GB on the Ada cards (more concurrent users, longer contexts) matters more than raw bandwidth. Both Ada cards also support FP8 paths in recent inference engines, which the A100 does not.
Image and video generation. Diffusion models are compute-heavy. The L40S and RTX 6000 Ada are the natural picks here, with 48 GB for larger models and higher resolutions.
Verdict: L40S or RTX 6000 Ada for most inference; A100 if your model fits in 40 GB and your engine is bandwidth-bound.
Fine-tuning
LoRA and QLoRA. All three handle LoRA on 7B to 13B models. The 48 GB cards allow larger batches or bigger base models in 4-bit.
Full fine-tuning and multi-GPU training. The A100 has the most mature training ecosystem, and A100 SXM4 machines are commonly built with several GPUs in one host. If you need a multi-GPU machine for data-parallel training, check whether A100 multi-GPU listings exist when you search. The Ada cards train well too, especially with BF16, but their lower memory bandwidth can slow large-batch training.
Verdict: pilot your own job. Run a few hundred steps on each listed option and compare cost per finished run (running price multiplied by hours), not hourly price.
Rendering
For Blender Cycles and other ray-traced work, the Ada cards win clearly: they have RT cores, which the A100 lacks entirely, and higher FP32 throughput. The RTX 6000 Ada is a workstation card designed for exactly this; the L40S is its datacenter cousin. The A100 is a poor choice for rendering.
Verdict: RTX 6000 Ada or L40S.
Summary table
| Workload | First choice | Alternative |
|---|---|---|
| LLM serving, model fits 40 GB, bandwidth-bound | A100 | L40S |
| LLM serving, larger model or many users | L40S or RTX 6000 Ada | A100 with a smaller model |
| Image and video generation | L40S or RTX 6000 Ada | A100 |
| LoRA fine-tuning | Any; pilot | Any |
| Multi-GPU training | A100 (if multi-GPU listed) | L40S multi-GPU |
| Rendering | RTX 6000 Ada or L40S | Not A100 |
Searching when your pick is not listed
GPU location on Cloudj is a country filter over a live list of host machines, and stock changes during the day. Open the L40S deploy page and set your countries. If your first choice is not listed:
- Clear the model and set a minimum VRAM of 40 GB. You will see all three (and others such as the RTX A6000) that are listed for your countries.
- Widen the country list, or remove it for batch jobs.
- Compare totals, including the bandwidth price, before deploying.
Regional searches are covered in L40S cloud in Europe, RTX 6000 Ada in Europe and A100 cloud in Asia. Live "from" prices are on each model page and on GPU cloud.
FAQ
Is the L40S better than the A100?
For inference with larger models, image generation and rendering, usually yes. For bandwidth-bound workloads that fit in 40 GB, the A100 can still be faster.
What is the difference between the L40S and the RTX 6000 Ada?
Same generation and similar VRAM. The L40S is a datacenter card; the RTX 6000 Ada is a workstation card. Choose on live price and availability.
Can I render with an A100?
It works in CUDA mode but has no RT cores. The Ada cards are much better for rendering.
Can a new account rent these cards?
Usually not straight away. All three need a higher account level; contact support from the console.
Compare them live
Open the L40S deploy page, set a 40 GB minimum and your countries, and see which of the three is listed now. All live prices are on GPU cloud.