Hopper
NVIDIA H100 80GB
wins
4
The NVIDIA H100 is the gold standard for AI training and inference. With 80GB HBM3 memory and NVLink 4.0 support, it dominates large-model training and multi-GPU clusters.
Ada Lovelace
NVIDIA L40S
wins
2
The L40S is NVIDIA's inference-optimized Ada Lovelace GPU. 48GB GDDR6 and 362 TFLOPS FP16 make it a strong choice for production inference at lower cost than H100.
Performance comparison
Visual bars — winner highlighted
NVIDIA H100 80GB
Metric
NVIDIA L40S
FP16 TFLOPS
Bandwidth
VRAM
TDP (lower=better)
TFLOPS/watt
GB/s per watt
TFLOPS/$
FP16 TFLOPS ÷ best on-demand $/hr
GB/s per $
Memory bandwidth ÷ best on-demand $/hr
Full specification table
| Spec | NVIDIA H100 80GB | NVIDIA L40S |
|---|---|---|
| Architecture | Hopper | Ada Lovelace |
| VRAM | 80 GB HBM3 | 48 GB GDDR6 |
| VRAM type | HBM3 | GDDR6 |
| Memory bandwidth | 3,350 GB/s | 864 GB/s |
| FP16 TFLOPS | 989 TFLOPS | 362 TFLOPS |
| TDP | 700 W | 350 W |
| TFLOPS/watt | 1.41 T/W | 1.03 T/W |
| GB/s per watt | 4.79 GB/s/W | 2.47 GB/s/W |
| NVLink | Yes | No |
| Release year | 2022 | 2023 |
| Best cloud price | $2.85/hr | $0.38/hr |
| TFLOPS/$ | 347.0 | 952.6 |
| GB/s per $ | 1175 | 2274 |
Live cloud pricing
On-demand pricing across providers — updated in real time
NVIDIA H100 80GB
NVIDIA L40S
Salad: Lowest priority tier — preemptible
Power efficiency analysis
TFLOPS/watt and GB/s/watt — critical for data center TCO
NVIDIA H100 80GB
NVIDIA L40S
TFLOPS/watt measures compute efficiency — how much AI throughput you get per watt of power consumed. For data centers with PUE of 1.2–1.5, a 10% improvement in TFLOPS/watt translates directly to lower electricity costs and cooling requirements. GB/s/watt measures memory bandwidth efficiency, which is the binding constraint for memory-bound LLM inference workloads.
Performance-per-dollar
TFLOPS/$ and GB/s/$ at current best cloud prices
NVIDIA H100 80GB
NVIDIA L40S
TFLOPS/$ = FP16 TFLOPS ÷ best on-demand $/hr. Higher is better — more compute per dollar. GB/s/$ = memory bandwidth ÷ best on-demand $/hr. For memory-bound inference workloads (most LLM serving), GB/s/$ is often a better predictor of cost-efficiency than TFLOPS/$. Prices are live from the Compute Comparison API and update in real time.
Workload Cost Comparison
Total job cost = best on-demand $/hr × GPU count × hours
NVIDIA H100 80GB costs 7.5× more at current best prices. NVIDIA H100 80GB has 803GB VRAM vs 486GB — larger models fit without offloading; 3.9× higher memory bandwidth (3350 vs 864 GB/s) improves token throughput for memory-bound inference.
| Workload | Scale | NVIDIA H100 80GB | NVIDIA L40S |
|---|---|---|---|
| 7B LLM fine-tune | 8× GPU, 24 h | $547 | $73▼ |
| 70B LLM fine-tune | 64× GPU, 72 h | $13k | $1.8k▼ |
| Inference server | 1× GPU, 730 h | $2.1k | $277▼ |
| Image gen batch | 4× GPU, 8 h | $91 | $12▼ |
Based on current best on-demand prices. Spot pricing typically 40–70% lower. Actual cost depends on provider, region, and availability.
When to choose each GPU
Choose NVIDIA H100 80GB for:
- LLM training
- Large-scale inference
- HPC workloads
Weaknesses / Limitations
- Highest rental cost of any widely available GPU
- SXM form factor requires NVLink-capable server chassis
- 700W TDP demands high-density power and cooling infrastructure
Choose NVIDIA L40S for:
- Inference
- Video AI
- Multi-modal workloads
Weaknesses / Limitations
- GDDR6 bandwidth (864 GB/s) is lower than HBM-based GPUs — bottleneck for memory-bound LLM serving
- No NVLink — cannot scale across cards for very large models
- Less suitable for large-scale training than H100 or A100
Popular GPU comparisons