NVIDIA L40S 48GB vs NVIDIA L20 48GB

The L40S delivers 3.1x the FP8 throughput of the L20 (733 vs 239 TFLOPS dense).

L40S: Ada Lovelace, 2023 L20: Ada Lovelace, 2024
FP8 dense lead
3.1x
L40S: 733 vs 239 TFLOPS
VRAM
48 vs 48 GB
Same capacity
Bandwidth
864 GB/s vs 864 GB/s
Even
TDP
350 W vs 275 W
L20 draws 75 W less

Specifications

L40S
L20
Architecture
Ada LovelaceAda Lovelace
Launch Year
20232024
Form Factor
PCIePCIe
VRAM
48 GB48 GB
Memory Bandwidth
864 GB/s864 GB/s
TDP
350 W275 W
Process Node
4nm

Performance (TFLOPS)

L40S
L20
FP64
1.4 TFlops
FP32
91.6 TFlops 59.8 TFlops
FP16
362.1 TFlops
733.0 TFLOPS sparse
119.5 TFlops
sparse not published
BF16
362.1 TFlops
733.0 TFLOPS sparse
119.5 TFlops
sparse not published
FP8
733.0 TFlops
1466.0 TFLOPS sparse
239.0 TFlops
sparse not published
INT8
733.0 TFlops
1466.0 TOPS sparse
239.0 TFlops
sparse not published

FLOPS by Precision

What actually differs

The L20 is the newer part: Ada Lovelace, launched in 2024, against the L40S's Ada Lovelace from 2023. Newer architectures typically add lower-precision formats and better throughput per watt, so check the precision rows your workload actually uses.

Dense throughput for the L40S against the L20: FP32 92 vs 60 TFLOPS, FP16 362 vs 120 TFLOPS, FP8 733 vs 239 TFLOPS. Sparse figures, where the vendor publishes them, appear under each dense number in the performance table.

Memory is 48 GB against 48 GB, fed at 864 GB/s versus 864 GB/s. More memory per GPU means larger models fit before you have to shard across cards, which often matters more than raw TFLOPS for inference.

Power budgets are 350 W for the L40S and 275 W for the L20. At FP32 that works out to 0.26 against 0.22 TFLOPS per watt.

For LLM training and inference, weight the FP8 and FP16 rows and memory capacity most heavily. For scientific and HPC workloads, the FP64 row is the one to read.

Frequently asked questions

Is the L40S faster than the L20?

At FP8 precision the L40S reaches 733 TFLOPS dense against 239 TFLOPS for the L20. The performance table on this page lists every published precision for both GPUs.

Which has more memory, the L40S or the L20?

Both GPUs carry 48 GB of VRAM. Memory bandwidth is 864 GB/s for the L40S and 864 GB/s for the L20.

How much power do the L40S and the L20 draw?

The L40S is rated at 350 W TDP and the L20 at 275 W. On FP32 throughput per watt, the L40S is the more efficient part.

Can I rent the L40S or the L20 in the cloud?

Yes. Live cloud listings tracked by Flopper start at $0.69 per GPU hour. The rental pricing section on this page lists current providers and rates for both GPUs.

Where to Rent

NVIDIA L40S 48GB

ProviderInstancePrice/GPU-hrChecked
DataCrunch
1L40S.20V (spot) $0.69 3h ago View →
Nebius
$0.74 3h ago View →
RunPod
L40S (Community Cloud) $0.79 3h ago View →
Cudo Compute
epyc-genoa-l40s-graphics $0.87 3h ago View →
Massed Compute
gpu_1x_l40s $0.88 3h ago View →
CoreWeave
nvidia-l40s (spot) $0.98 3h ago View →
RunPod
L40S (Secure Cloud) $0.99 3h ago View →
DataCrunch
1L40S.20V (on-demand) $1.37 3h ago View →
Crusoe
$1.50 3h ago View →
Nebius
$1.55 3h ago View →
DigitalOcean
NVIDIA L40S $1.57 3h ago View →
Vultr
vcg-l40s-16c-180g-48vram $1.67 3h ago View →
OVHcloud
l40s-360 $1.80 3h ago View →
CoreWeave
nvidia-l40s (on-demand) $2.25 3h ago View →
All L40S listings and price history →

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.