GPU comparison
L40S vs RTX 5090.
Pick the L40S for memory: 48 GB vs 32 GB. A QLoRA fine-tune of a 70B model (about 46 GB) fits on one L40S, not on one RTX 5090. The RTX 5090 has 2.1x the memory bandwidth. For models under 32 GB pick by price.
Rent either on Runcrate
- L40Sfrom $1.26/hr
$0.026 per GB/hr
- RTX 5090Cheaperfrom $0.72/hr per GPU
on an 8-GPU machine
$0.022 per GB/hr
Live hourly price, then per GB of memory.
Which is better for
Which to pick for each job. Memory fit, from the datasheets.
Hold a 70B model in 8-bit weights
70 GB of 8-bit weights
L40S: 2 GPUs (96 GB)
RTX 5090: 3 GPUs, more than one 2x node
L40S. Fits in one node (2 GPUs).
Serve an 8B model in 16-bit weights
16 GB of 16-bit weights
L40S: 1 GPU, 32 GB spare
RTX 5090: 1 GPU, 16 GB spare
RTX 5090. Same GPU count, 2.1x the memory bandwidth.
Fine-tune a 70B model with QLoRA
About 46 GB with 4-bit weights
L40S: 1 GPU, 2 GB spare
RTX 5090: 2 GPUs (64 GB)
L40S. Fits on one GPU.
Compare another pair
Ada Lovelace
NVIDIA L40S
- Memory
- 48 GB GDDR6
- Bandwidth
- 864 GB/s
Blackwell
NVIDIA RTX 5090
on an 8-GPU machine
- Memory
- 32 GB GDDR7
- Bandwidth
- 1.8 TB/s
At a glance
Spec by spec. The longer bar is the better number.
Full specs
The complete spec sheet.
| Spec | NVIDIA L40S | NVIDIA RTX 5090 |
|---|---|---|
| Architecture | Ada Lovelace | Blackwell |
| Memory | 48 GB | 32 GB |
| Memory type | GDDR6 | GDDR7 |
| Memory bandwidth | 864 GB/s | 1.8 TB/s |
| FP32 | 91.6 TFLOPS | 104.8 TFLOPS |
| FP16 | 362 TFLOPS, tensor | 209.5 TFLOPS, non-tensor |
| INT8 | 733 TOPS | 838 TOPS |
| TDP | 350 W | 575 W |
| Form factor | PCIe | Consumer |
| NVLink | No | No |
| GPUs per node | 4 | 2 |
| Hourly price (live) | from $1.26/hr | from $0.72/hr per GPU on an 8-GPU machine |
Questions
L40S or RTX 5090. What buyers ask first.
Other comparisons