A100 80GB SXM vs L40S
Side-by-side comparison of the NVIDIA A100 80GB SXM and the NVIDIA L40S for AI inference workloads.
Specifications
| Spec | A100 80GB SXM | L40S |
|---|---|---|
| Generation | ampere | ada |
| Memory Type | HBM2e | GDDR6 |
| VRAM | 80 GB | 48 GB |
| Memory Bandwidth | 2,039 GB/s | 864 GB/s |
| BF16 TFLOPS | 312 | 362 |
| FP16 TFLOPS | 312 | 362 |
| FP8 TFLOPS | 312 | 733 |
| INT8 TOPS | 624 | 733 |
| TDP | 400 W | 350 W |
| Interconnect | nvlink | pcie |
| NVLink Bandwidth | 600 GB/s | N/A |
| Max GPUs per Node | 8 | 8 |
| PCIe Gen | Gen 4 | Gen 4 |
| CUDA Compute Capability | 8 | 8.9 |
Pricing
A100 80GB SXM
| Provider | On-Demand | Reserved | Spot |
|---|---|---|---|
| runpod | $2.72/hr | - | $2.09/hr |
| lambda | $1.99/hr | $1.49/hr | - |
| coreweave | $2.21/hr | $1.62/hr | - |
| aws | $3.67/hr | $2.39/hr | - |
| gcp | $3.67/hr | $2.48/hr | - |
| azure | $3.67/hr | $2.45/hr | - |
| vast ai | $1.80/hr | - | $1.30/hr |
| tensordock | $1.79/hr | - | $1.29/hr |
| fluidstack | $1.69/hr | - | $1.19/hr |
L40S
| Provider | On-Demand | Reserved | Spot |
|---|---|---|---|
| runpod | $1.90/hr | - | $1.49/hr |
| lambda | $1.59/hr | $1.19/hr | - |
| coreweave | $1.84/hr | $1.34/hr | - |
| aws | $2.56/hr | $1.69/hr | - |
| gcp | $2.45/hr | $1.62/hr | - |
| vast ai | $1.29/hr | - | $0.95/hr |
| tensordock | $1.19/hr | - | $0.89/hr |
| fluidstack | $1.09/hr | - | $0.85/hr |
Cheapest available rate: A100 80GB SXM at $1.19/hr vs L40S at $0.85/hr — L40S is +40% cheaper
Efficiency Metrics
TFLOPS / Watt
0.8
A100 80GB SXM
1.0
L40S
BF16
VRAM / Dollar
67.2
A100 80GB SXM
56.5
L40S
GB/$/hr
Bandwidth / Watt
5.1
A100 80GB SXM
2.5
L40S
GB/s/W
Models (FP16, 1 GPU)
182.0
A100 80GB SXM
162.0
L40S
Model Compatibility (FP16, Single GPU)
Only on A100 80GB SXM (20)
- Yi 1.5 34B
- Qwen 2.5 32B
- Qwen 2.5 Coder 32B
- Aya 23 35B
- Command R
- DeepSeek Coder 33B
- DeepSeek R1 Distill 32B
- Gemma 2 27B
- Gemma 3 27B
- InternVL2 26B
- Vicuna 33B
- Code Llama 34B
- Mistral Small 24B
- Mistral Small 3.1 24B
- Qwen 3 32B
- WizardCoder 33B
- Qwen 3 30B-A3B
- JAIS 30B
- Command R (August 2024)
- MPT 30B
Both (162)
- Yi 1.5 9B
- Yi Coder 9B
- GTE Qwen2 7B
- Marco O1
- Qwen 1.5 MoE A2.7B
- Qwen 2 Audio 7B
- Qwen 2.5 14B
- Qwen 2.5 3B
- OLMo 2 13B
- OLMo 2 7B
- Amazon Nova Lite
- OpenELM 3B
- BGE Large EN v1.5
- BGE M3
- Baichuan 2 13B
- OctoCoder 15B
- StarCoder2 15B
- StarCoder2 3B
- StarCoder2 7B
- Aya 23 8B
- +142 more
Only on L40S (0)
None
Summary
The A100 80GB SXM (ampere generation) offers 80GB of HBM2e with 312 BF16 TFLOPS and 2,039 GB/s memory bandwidth at 400W TDP.
The L40S (ada generation) offers 48GB of GDDR6 with 362 BF16 TFLOPS and 864 GB/s memory bandwidth at 350W TDP.
The A100 80GB SXM has +67% more VRAM, allowing it to run larger models without multi-GPU setups.
From a cost perspective, the L40S is more affordable at $0.85/hr vs $1.19/hr for the A100 80GB SXM.