What can I run on L40?
48 GB VRAMAda Lovelace300WView matrix for this target →
Not the L40S (different SKU, lower power). NVIDIA's L40 page does not show a Tensor TFLOPS headline in its spec table. Not listed as a Hugging Face Jobs flavor.
Policy by policy
| Policy | Params | Inference memory (sourced, with precision) | On L40 | Best measured |
|---|---|---|---|---|
| GR00T N1.7 VLA | 3B | ≥16 GB | measured | 28.57 Hz · TensorRT · bf16 · 2 cam |
| ACT Imitation | 80M | 211 MB measured fp32 | no data | |
| Diffusion Policy Imitation | — | 1.12 GB measured fp32 | no data | |
| EO-1 VLA | 3B | ≥6.5 GB | no data | |
| Evo-1 VLA | 770M | — | no data | |
| Fast-WAM World model | 6B | — | no data | |
| HIL-SERL RL | — | — | no data | |
| LingBot-VA World model | 5.3B | ≥18 GB | no data | |
| MolmoAct2 VLA | — | 12.1 GB measured bf16 | no data | |
| Multitask DiT Imitation | 450M | — | no data | |
| OpenVLA VLA | 7.5B | 15 GB measured bf16 | no data | |
| OpenVLA-OFT VLA | — | ≥16 GB | no data | |
| SmolVLA VLA | 450M | 1.75 GB measured fp32 | no data | |
| TD-MPC RL | — | — | no data | |
| VLA-JEPA VLA | — | — | no data | |
| VQ-BeT Imitation | — | — | no data | |
| WALL-OSS VLA | — | — | no data | |
| X-VLA VLA | 900M | — | no data | |
| π0 VLA | 3.3B | 13.32 GB measured fp32 | no data | |
| π0-FAST VLA | — | ≥8 GB | no data | |
| π0.5 VLA | — | ≥8 GB | no data |
“Exceeds memory” compares a policy's upstream-stated minimum inference memory (never a measured peak, which is precision-specific) with this target's VRAM. It is arithmetic, not observation, and is labelled inferred. Quantised or split runtimes can change the answer; a measured run always wins.
All measured runs on L40
| Policy | Runtime | Precision | Cams | p50 / p95 ms | Inference | Platform | Source |
|---|---|---|---|---|---|---|---|
| GR00T N1.7 | TensorRT | bf16 | 2 @256x256 | 35 / — | 28.57 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16). | T3 · community reported |
| GR00T N1.7 | torch.compile | bf16 | 2 @256x256 | 66.8 / — | 14.97 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16). | T3 · community reported |
| GR00T N1.7 | PyTorch eager | bf16 | 2 @256x256 | 123.7 / — | 8.08 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16). | T3 · community reported |
Sources
- T1 · upstream documentedvendor-docsverified 2026-10-24 (1d ago)
L40 spec table: Ada Lovelace, 48 GB GDDR6 with ECC, 300W max power.
“GPU Architecture NVIDIA Ada Lovelace Architecture GPU Memory Bandwidth 864GB/s GPU Memory 48 GB GDDR6 with ECC Display Connectors 4 x DP 1.4a Max Power Consumption 300W”
https://www.nvidia.com/en-us/data-center/l40/