Ada Lovelace
NVIDIA L40S
wins
4
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.
Ada Lovelace
NVIDIA RTX 4090
wins
2
The RTX 4090 is the most cost-effective consumer GPU for AI workloads. At $0.40–$0.80/hr on GPU clouds, it's ideal for fine-tuning and small-model inference.
Performance comparison
Visual bars — winner highlighted
NVIDIA L40S
Metric
NVIDIA RTX 4090
FP16 TFLOPS
Bandwidth
VRAM
TDP (lower=better)
TFLOPS/watt
GB/s per watt
TFLOPS/$
FP16 TFLOPS ÷ best on-demand $/hr
GB/s per $
Memory bandwidth ÷ best on-demand $/hr
Full specification table
| Spec | NVIDIA L40S | NVIDIA RTX 4090 |
|---|---|---|
| Architecture | Ada Lovelace | Ada Lovelace |
| VRAM | 48 GB GDDR6 | 24 GB GDDR6X |
| VRAM type | GDDR6 | GDDR6X |
| Memory bandwidth | 864 GB/s | 1,008 GB/s |
| FP16 TFLOPS | 362 TFLOPS | 165 TFLOPS |
| TDP | 350 W | 450 W |
| TFLOPS/watt | 1.03 T/W | 0.37 T/W |
| GB/s per watt | 2.47 GB/s/W | 2.24 GB/s/W |
| NVLink | No | No |
| Release year | 2023 | 2022 |
| Best cloud price | $0.38/hr | $0.13/hr |
| TFLOPS/$ | 952.6 | 1240.6 |
| GB/s per $ | 2274 | 7579 |
Live cloud pricing
On-demand pricing across providers — updated in real time
NVIDIA L40S
NVIDIA RTX 4090
Salad: Lowest priority tier — preemptible
Power efficiency analysis
TFLOPS/watt and GB/s/watt — critical for data center TCO
NVIDIA L40S
NVIDIA RTX 4090
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 L40S
NVIDIA RTX 4090
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 L40S costs 2.9× more at current best prices. NVIDIA L40S has 486GB VRAM vs 246GB — larger models fit without offloading; 2.2× the FP16 compute (362 vs 165 TFLOPS) supports faster training and higher batch throughput.
| Workload | Scale | NVIDIA L40S | NVIDIA RTX 4090 |
|---|---|---|---|
| 7B LLM fine-tune | 8× GPU, 24 h | $73 | $26▼ |
| 70B LLM fine-tune | 64× GPU, 72 h | $1.8k | $613▼ |
| Inference server | 1× GPU, 730 h | $277 | $97▼ |
| Image gen batch | 4× GPU, 8 h | $12 | $4▼ |
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 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
Choose NVIDIA RTX 4090 for:
- Fine-tuning
- Small model inference
- Cost-effective compute
Weaknesses / Limitations
- 24 GB VRAM limits models to ~34B at INT4 — cannot run 70B+ unquantized
- No NVLink — cannot pool VRAM across cards
- Consumer GPU — no ECC memory, lower reliability for 24/7 production
Popular GPU comparisons