New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
NVIDIA L40S 48GB
Runs downloadable LLMs up to ~71B at Q4, ~140.8-167.8 tok/s on an 8B model.
What you can run
At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to system RAM (CPU/PCIe offload) and drops to the degraded figure.
| Model size | FP16 | Q8 | Q4 |
|---|---|---|---|
| 3B |
105.1-125.3 t/s ≤559k ctx · then ~4.9-5.8 (slow) |
210.2-250.6 t/s ≤605k ctx · then ~9.7-11.6 (slow) |
375.4-447.4 t/s ≤625k ctx · then ~17.4-20.7 (slow) |
| 7-8B |
39.4-47 t/s ≤203k ctx · then ~1.8-2.2 (slow) |
78.8-94 t/s ≤264k ctx · then ~3.7-4.4 (slow) |
140.8-167.8 t/s ≤291k ctx · then ~6.5-7.8 (slow) |
| 13B |
24.3-28.9 t/s ≤102k ctx · then ~1.1-1.3 (slow) |
48.5-57.8 t/s ≤181k ctx · then ~2.2-2.7 (slow) |
86.6-103.3 t/s ≤216k ctx · then ~4-4.8 (slow) |
| 32B | won't run |
19.7-23.5 t/s ≤41k ctx · then ~0.9-1.1 (slow) |
35.2-41.9 t/s ≤94k ctx · then ~1.6-1.9 (slow) |
| 70B | won't run | won't run |
16.1-19.2 t/s ≤11k ctx · then ~0.7-0.9 (slow) |
| 120B | won't run | won't run | won't run |
| 400B | won't run | won't run | won't run |
Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (NVIDIA). Cyan = fits ~44.2 GB usable. Beyond "ctx", KV spills to system RAM (CPU/PCIe offload) (rough degraded speed shown).
What else you need for a full build▶
Minimum supporting components for a desktop built around the NVIDIA L40S 48GB. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.
Run popular models on the NVIDIA L40S 48GB
More hardware like the NVIDIA L40S 48GB
About the NVIDIA L40S 48GB
NVIDIA L40S 48GB is a discrete GPU from NVIDIA on Ada Lovelace, 2023. For local genAI the numbers that matter are its 48 GB (how big a model + context fits) and 864 GB/s bandwidth (generation speed). Plan for a ~750W power supply.
Frequently asked questions
What size LLM can the NVIDIA L40S 48GB run?▶
With ~44.2 GB usable it runs up to ~71B at Q4 (~20B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.
How many tokens/sec does the NVIDIA L40S 48GB generate?▶
Generation is bandwidth-bound. At 864 GB/s an 8B model at Q4 runs ~140.8-167.8 tok/s (estimate).
Why does long context slow it down?▶
The KV cache for the context must also fit fast memory; once weights+KV exceed 44.2 GB it spills to system RAM (CPU/PCIe offload) and throughput collapses. Each row shows the max usable context.
What else do I need to run the NVIDIA L40S 48GB?▶
Besides the card: a ≥750W 80+ Gold power supply, at least 48GB of system RAM (2× the 48GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.