Runs easilyfp16 · 4K ctx · batch 1

Run Llama 3.1 8B on L40S

At fp16 precision, Llama 3.1 8B (8.0 B parameters) needs roughly 19.5 GB of VRAM. L40S has 48 GB, leaving 28.5 GB of headroom — plenty of room for longer contexts and larger batch sizes. Expected throughput ≈ 112 tokens/sec on a single card.

41%
VRAM
Runs easily
Estimated VRAM (fp16)
19.5 GB
of 48 GB on L40S· 28.5 GB free
Tok/sec
112
TFTT
597ms
Power
134W
Rent on GPUniq marketplace
L40S
Live prices from verified providers · billed hourly, no commitment.
Rent now

Memory breakdown (fp16)

Model Weights
16.00 GB
KV Cache
0.50 GB
Activations
2.00 GB
Framework Overhead
1.04 GB

Quantization comparison

Lower precision = less VRAM with a small quality trade-off. Quality order: FP16 > INT8 > INT4.

PrecisionVRAMUtilisationFits on L40S?
FP16 (full precision)19.5 GB41%Runs easily
INT8 (8-bit)10.5 GB22%Runs easily
INT4 / Q4 (4-bit)6.0 GB13%Runs easily

Frequently asked

Can L40S run Llama 3.1 8B?
Yes. Llama 3.1 8B needs ≈ 19.5 GB VRAM at fp16 and L40S provides 48 GB. Expected throughput is 112 tokens/sec per GPU.
How much VRAM does Llama 3.1 8B use?
About 19.5 GB at fp16, 6.0 GB at INT4 (for a 4K context, batch size 1). Longer contexts add to KV cache size; larger batches increase activations.
What's the fastest way to run Llama 3.1 8B on L40S?
Use a production inference engine like vLLM, SGLang, or TensorRT-LLM with paged attention — they cut KV cache 2–3× vs the naive estimate and batch multiple requests efficiently. For L40S, enable Flash Attention 2.
Where can I rent a L40S?
GPUniq aggregates live L40S offers from verified providers. You can deploy an instance in about a minute and pay hourly with no commitment.
Try it live

Chat with Llama 3.1 8B — no setup

Send a prompt and see how Llama 3.1 8B responds directly in our chat. No installation, no GPU required to test.

Open in chat
Want to tweak sequence length, batch size, or fine-tuning? Open the full calculator →