Short answer: NVIDIA A40 and L40S are both passive dual-slot 48GB GDDR6 ECC data-center cards, but they are not interchangeable. A40 is a 300W Ampere card with compute capability 8.6 and two-card NVLink support; L40S is a 350W Ada card with compute capability 8.9, higher published bandwidth and newer media engines, but no NVLink. Both require validated server airflow.
A40 vs L40S specifications
| Specification | NVIDIA A40 | NVIDIA L40S |
|---|---|---|
| Architecture / compute capability | Ampere / 8.6 | Ada / 8.9 |
| Memory | 48GB GDDR6 ECC | 48GB GDDR6 ECC |
| Published bandwidth | 696GB/s | 864GB/s |
| Maximum power | 300W | 350W |
| Cooling / size | Passive, 10.5-inch dual slot | Passive, 10.5-inch dual slot |
| Power connector | CPU 8-pin | 16-pin |
| NVLink | Two-card bridge, 112.5GB/s bidirectional | No |
| MIG | No | No |
| NVENC / NVDEC | 1 / 2, including AV1 decode | 3 / 3, including AV1 encode/decode |
Model fit is the same 48GB calculation
Neither generation changes the nominal memory ceiling. Count the exact model artifact or tensors, KV cache, runtime workspaces, context, batch size and concurrency. A 70B-class 4-bit artifact may fit in some 48GB configurations but remain tight after cache and overhead.
Use the 70B VRAM guide and LLM VRAM calculator. If 48GB is not enough, compare the 48GB and 80GB tiers before adding a second card.
Both cards need server airflow
A40 and L40S use passive heatsinks. The host must force appropriate air through the card across the expected ambient temperature and sustained workload. Their dual-slot dimensions do not make them workstation cards. Validate airflow direction, pressure, inlet temperature, adjacent-card spacing and sustained temperatures in the final chassis.
Power integration differs
- A40: NVIDIA documents a 300W maximum and CPU 8-pin connector.
- L40S: NVIDIA documents 350W maximum power and a 16-pin connector.
- For either: verify the exact board documentation and supplied cable. Connector appearance is not proof of a compatible pinout.
Software-generation tradeoff
A40 targets compute capability 8.6; L40S targets 8.9. Confirm the driver, CUDA toolkit, framework build, runtime, quantization kernels and compiled extensions for the exact target. A newer architecture can require a newer binary even when both cards appear under a generic “CUDA supported” label.
The A40 guide covers its display and used-card integration details. The L40S guide compares it with the active workstation RTX 6000 Ada.
Does L40S’s bandwidth prove it is faster?
No. The 864GB/s and 696GB/s figures are published peak memory bandwidth. Measured local-AI performance also depends on architecture, kernels, precision, runtime, prompt processing, context, batch size, CPU and system configuration. Use the same reproducible workload on each accessible system.
NVLink versus newer media engines
A40 supports a two-card NVLink bridge; L40S does not. NVLink still requires application support to scale memory or performance and does not create one automatic 96GB device. L40S instead provides three NVENC and three NVDEC engines with AV1 encode/decode support, which can matter for media-heavy pipelines but not for LLM memory capacity.
For multi-card model placement, read the multi-GPU VRAM guide.
Choose A40 when
- Your verified runtime supports compute capability 8.6.
- The server provides appropriate 300W power, CPU 8-pin cabling and forced airflow.
- A supported two-card NVLink configuration has application value.
- Used-card evidence and complete-system cost make the older path defensible.
Choose L40S when
- Your software stack supports compute capability 8.9 and benefits from the Ada generation.
- The server is qualified for 350W, 16-pin power and passive cooling.
- NVLink is not required.
- Newer media engines or the L40S platform ecosystem matter to the workload.
Official NVIDIA sources
- NVIDIA A40 datasheet
- NVIDIA L40S specifications
- NVIDIA CUDA compute-capability table
- NVIDIA NVLink bridge compatibility and application-support caveat
Reviewed October 25, 2026. Specifications are sourced from NVIDIA. HighVRAM has not independently benchmarked A40 and L40S under one controlled workload. Availability, used condition, licensing and prices change.
Does 80GB matter more than the 48GB generation change? Compare L40S with A100 40GB/80GB using exact Qwen3.5 boundaries and server-platform safeguards.
Does 80GB matter more than the generation change? Compare A40 with A100 40GB/80GB before moving from Ampere to L40S.
What if the Ada alternative is a workstation card? Read A40 versus RTX 6000 Ada to hold 48GB and 300W constant while changing cooling, connector, NVLink and display role.