Ada Lovelace
NVIDIA L40
wins
4
The NVIDIA L40 is the predecessor to the L40S, offering 48GB GDDR6 and Ada Lovelace architecture.
Ampere
NVIDIA RTX 3090
wins
1
The RTX 3090 remains a popular budget option for AI workloads with 24GB GDDR6X.
Performance comparison
Visual bars — winner highlighted
NVIDIA L40
Metric
NVIDIA RTX 3090
FP16 TFLOPS
Bandwidth
VRAM
TDP (lower=better)
TFLOPS/watt
GB/s per watt
Full specification table
| Spec | NVIDIA L40 | NVIDIA RTX 3090 |
|---|---|---|
| Architecture | Ada Lovelace | Ampere |
| VRAM | 48 GB GDDR6 | 24 GB GDDR6X |
| VRAM type | GDDR6 | GDDR6X |
| Memory bandwidth | 864 GB/s | 936 GB/s |
| FP16 TFLOPS | 181 TFLOPS | 71 TFLOPS |
| TDP | 300 W | 350 W |
| TFLOPS/watt | 0.60 T/W | 0.20 T/W |
| GB/s per watt | 2.88 GB/s/W | 2.67 GB/s/W |
| NVLink | No | No |
| Release year | 2022 | 2020 |
Live cloud pricing
On-demand pricing across providers — updated in real time
NVIDIA L40
NVIDIA RTX 3090
No live pricing data available.
Power efficiency analysis
TFLOPS/watt and GB/s/watt — critical for data center TCO
NVIDIA L40
NVIDIA RTX 3090
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 L40
NVIDIA RTX 3090
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
| Workload | Scale | NVIDIA L40 | NVIDIA RTX 3090 |
|---|---|---|---|
| 7B LLM fine-tune | 8× GPU, 24 h | $180 | — |
| 70B LLM fine-tune | 64× GPU, 72 h | $4.3k | — |
| Inference server | 1× GPU, 730 h | $684 | — |
| Image gen batch | 4× GPU, 8 h | $30 | — |
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 L40 for:
- Professional inference
- Rendering
- Visualization AI
Weaknesses / Limitations
- Lower FP16 TFLOPS (181) than L40S (362) at the same VRAM and bandwidth
- Superseded by L40S — less available on new deployments
- GDDR6 bandwidth lower than HBM-based data center GPUs
Choose NVIDIA RTX 3090 for:
- Budget fine-tuning
- Small model inference
- Hobbyist AI
Weaknesses / Limitations
- Older Ampere architecture — significantly lower TFLOPS than Ada equivalents
- 350W TDP — high power draw for a 24 GB consumer card
- No ECC memory — not suitable for production reliability requirements
Popular GPU comparisons