Quick answer
L40S vs A100 for inference and fine-tuning?
L40S vs A100 for inference and fine-tuning: pick the L40S for FP8 inference and image or video generation on models that fit in 48 GB. Pick the A100 80GB for fine-tuning, models between 48 and 80 GB, and multi-GPU jobs: it has NVLink and up to 2.4 times the bandwidth. As of 29 October 2026, GPUwithCrypto rents the L40S at $0.82/hour and the A100 80GB at $0.92/hour, paid in crypto with no KYC.
How do the L40S and A100 specs compare?
Official NVIDIA figures, read on 29 October 2026. NVIDIA prints both dense and sparse Tensor values for these cards; the table shows dense peak TFLOPS. Peak figures are not benchmarks.
| Spec | L40S | A100 80GB SXM | A100 80GB PCIe |
|---|---|---|---|
| Architecture | Ada Lovelace | Ampere | Ampere |
| Memory | 48 GB GDDR6 | 80 GB HBM2e | 80 GB HBM2e |
| Memory bandwidth | 864 GB/s | 2,039 GB/s | 1,935 GB/s |
| FP16 / BF16 Tensor, dense | 362.05 TFLOPS | 312 TFLOPS | 312 TFLOPS |
| FP8 Tensor, dense | 733 TFLOPS | N/A | N/A |
| NVLink | No | 600 GB/s | 600 GB/s (2-GPU bridge) |
| PCIe | Gen4 x16 | Gen4 | Gen4 |
| Max power | 350 W | 400 W | 300 W |
| Launch | 2023 | 2020 | 2021 |
Sparse figures: L40S 733 TFLOPS BF16 and 1,466 TFLOPS FP8; A100 624 TFLOPS BF16. NVIDIA lists FP8 as not available on Ampere. The L40S page states 'NVLink Support: No'.
Is the L40S faster than the A100?
For math, slightly: the L40S has 362.05 dense BF16 TFLOPS against 312 (1.16 times), plus FP8 at 733 TFLOPS, which the A100 lacks. For memory, no: the A100 80GB SXM moves 2,039 GB/s, 2.36 times the L40S's 864 GB/s (2.24 times for the A100 PCIe). These are NVIDIA's peak figures, not benchmarks.
- Token-by-token LLM decoding depends mostly on bandwidth, where the A100 leads.
- Image and video diffusion and large-batch prefill are compute-heavy, which favors the L40S's higher peak and FP8.
- Multi-GPU work favors the A100. The L40S has no NVLink, so its GPUs talk over PCIe Gen4.
Instances have 1 to 8 GPUs depending on the host. Before you start, the dashboard shows each host’s country, GPU count, form factor (SXM or PCIe) and interconnect.
Which models fit in 48 GB, and which need 80 GB?
Rule of thumb for inference: memory ≈ parameters (billions) × bytes per parameter × 1.2, with 2 bytes for BF16, 1 for FP8 and 0.5 for 4-bit (EleutherAI). LoRA planning = 2 bytes per parameter × 1.2. QLoRA ≈ 0.75 GB per billion parameters. Formula and more models: VRAM guide.
| Model and task | Memory needed | L40S 48GB | A100 80GB |
|---|---|---|---|
| Mistral 7B, BF16 inference | 17.5 GB | 1 GPU | 1 GPU |
| Llama 3.1 8B, BF16 inference | 19.2 GB | 1 GPU | 1 GPU |
| Llama 3.1 8B, LoRA fine-tune | 19.2 GB | 1 GPU | 1 GPU |
| FLUX.1 [dev], BF16 weights | 24 GB + text encoders | 1 GPU | 1 GPU |
| FLUX.1 [dev], LoRA fine-tune | 28.8 GB | 1 GPU | 1 GPU |
| Llama 3.1 70B, 4-bit inference | 42 GB | 1 GPU | 1 GPU |
| Qwen2.5 72B, 4-bit inference | 43.6 GB | 1 GPU | 1 GPU |
| Llama 3.1 70B, QLoRA fine-tune | 52.5 GB | 2 GPUs (borderline on 1) | 1 GPU |
| Llama 3.1 70B, FP8 inference | 84 GB | 2 GPUs, no NVLink | 2 GPUs, no FP8 speed-up |
| Llama 3.1 70B, LoRA fine-tune | 168 GB | 4 GPUs | 3 GPUs |
Naive fit = the number of cards whose total memory holds the estimate; it ignores multi-GPU overhead and memory reserved by the driver. QLoRA's authors report a 65B model fine-tuned on one 48 GB GPU, hence 'borderline'. FLUX.1 also loads T5-XXL, CLIP-L and a VAE, which can be offloaded to CPU between steps.
How much do the L40S and A100 cost per hour?
Per GPU-hour, USD. 'Lowest' and 'median' are self-serve on-demand list prices from GPUwithCrypto's price check of 29 October 2026; spot and marketplace prices are excluded. Vast.ai is a marketplace, shown apart. All GPUs: pricing.
| GPU | GPUwithCrypto | Lowest on-demand elsewhere | Median on-demand | Providers checked | Vast.ai lowest offer |
|---|---|---|---|---|---|
| L40S 48GB | $0.82/hour | $0.97 (Massed Compute) | $1.67 | 11 | $0.33 |
| A100 80GB | $0.92/hour | $1.09 (Thunder Compute) | $2.23 | 13 | $0.48 (SXM), $0.34 (PCIe) |
Prices checked on
At the lowest prices the A100 80GB costs 12% more per hour than the L40S; at the median, about a third more. Vast.ai figures: Vast.ai's public price feed, cheapest rentable offer on 29 October 2026. Taxes excluded.
Which is cheaper per GB of VRAM and per TFLOPS?
The A100 is cheaper per GB, the L40S per TFLOPS. According to GPUwithCrypto's price check of 29 October 2026, computed as hourly price ÷ GB of memory:
- A100 80GB: 1.36 cents per GB-hour at the lowest on-demand price ($1.09 ÷ 80), 2.79 cents at the median ($2.23 ÷ 80).
- L40S 48GB: 2.02 cents per GB-hour at the lowest price ($0.97 ÷ 48), 3.48 cents at the median ($1.67 ÷ 48).
A GB of A100 memory is 33% cheaper per hour at the lowest prices and 20% cheaper at the median. At GPUwithCrypto the same division is $0.012 and $0.017.
Per dollar at the lowest prices, the L40S gives 30% more peak BF16 math (373 vs 286 TFLOPS per $/hour) and 756 FP8 TFLOPS per $/hour. The A100 SXM gives 2.1 times the bandwidth per dollar (1,871 vs 891 GB/s per $/hour).
What does one job cost on an L40S vs an A100?
Worked examples (assumptions, not benchmarks):
- Serving Llama 3.1 8B for a month (720 hours). It fits one card of either type. The L40S is the cheaper choice if its throughput is enough for your traffic.
- A QLoRA fine-tune of Llama 3.1 70B planned at 24 hours. It needs about 52.5 GB: one A100, or two L40S. Assuming the same 24 hours on both setups, the A100 costs 44% less at the lowest prices and 33% less at the median.
Cost of the two jobs
| Price basis | 8B serving, 720 h: L40S | 8B serving, 720 h: A100 | 70B QLoRA, 24 h: 2 × L40S | 70B QLoRA, 24 h: 1 × A100 |
|---|---|---|---|---|
| GPUwithCrypto | $590.40 | $662.40 | 2 × $19.68 | $22.08 |
| Lowest on-demand elsewhere | $698.40 | $784.80 | $46.56 | $26.16 |
| Median on-demand | $1,202.40 | $1,605.60 | $80.16 | $53.52 |
Prices checked on
Hours are assumptions. GPU time only.
Which should I rent for inference, and which for fine-tuning?
For inference, rent the L40S when the model fits in 48 GB with room for its KV cache: 7B–8B models in BF16, 70B models in 4-bit, and image or video generation with FLUX or SDXL. FP8 checkpoints run on its FP8 Tensor Cores. Rent the A100 80GB when the model needs 48–80 GB, when long contexts or large batches fill the cache, or when decoding speed is limited by bandwidth.
For fine-tuning, the A100 80GB is the safer default. QLoRA on a 70B model fits one card, and NVLink at 600 GB/s helps multi-GPU runs. LoRA on 7B–8B models and FLUX LoRAs fit either card.
More: best GPU for LLM inference, best GPU for ComfyUI and Flux, best GPU to fine-tune Llama 70B.
Questions about the L40S and the A100
Is the L40S better than the A100 for inference?
For models that fit in 48 GB and for FP8 or image generation, yes: it has 1.16 times the dense BF16 peak and FP8 Tensor Cores. For larger models or bandwidth-bound decoding, the A100 80GB leads with 80 GB and 2.36 times the bandwidth (SXM). These are NVIDIA peak figures, not benchmarks.
Is the L40S good for fine-tuning?
For LoRA on 7B–8B models and FLUX LoRAs, yes: they fit in 48 GB. QLoRA on a 70B model needs about 52.5 GB by our rule, borderline on one L40S. Multi-GPU fine-tuning runs over PCIe, because the L40S has no NVLink.
How much does an L40S cost per hour compared with an A100?
On 29 October 2026 the lowest self-serve on-demand prices we found were $0.97 for an L40S and $1.09 for an A100 80GB; the medians were $1.67 and $2.23. GPUwithCrypto charges $0.82/hour and $0.92/hour.
Does the L40S support NVLink?
No. NVIDIA's L40S page states 'NVLink Support: No'; GPUs talk over PCIe Gen4 x16 at 64 GB/s bidirectional. The A100 80GB has NVLink at 600 GB/s.
Can I rent an L40S or an A100 with crypto?
Yes. GPUwithCrypto rents both for BTC, USDT, USDC, ETH, SOL, LTC, with no ID and no card. Of the 11 providers with an on-demand L40S price on 29 October 2026, 5 accept crypto: Massed Compute, RunPod (KYC first), Paperspace/DigitalOcean, Vultr and Spheron.
Rent an L40S or an A100 with crypto
L40S $0.82/hour · A100 80GB $0.92/hour · billed per minute · no KYC.
Sources (10)
- NVIDIA L40S GPU — NVIDIA (read 2026-10-29)
- NVIDIA A100 Tensor Core GPU — NVIDIA (read 2026-10-29)
- NVIDIA A100 datasheet — NVIDIA (read 2026-10-29)
- Transformer Math 101 — EleutherAI (read 2026-10-29)
- QLoRA: Efficient Finetuning of Quantized LLMs — arXiv (read 2026-10-29)
- LoRA: Low-Rank Adaptation of Large Language Models — arXiv (read 2026-10-29)
- FLUX.1 [dev] model card — Black Forest Labs / Hugging Face (read 2026-10-29)
- Massed Compute pricing vm.massedcompute.com/pricing — Massed Compute (read 2026-10-29)
- Thunder Compute pricing thundercompute.com/pricing — Thunder Compute (read 2026-10-29)
- Vast.ai public GPU pricing feed — Vast.ai (read 2026-10-29)