Can L40S 48GB run Llama-3.3-70B-Instruct?

NVIDIA · 48 GB GDDR6 · 864 GB/s bandwidth · Meta · 70.6B · model ctx up to 128K

No — Llama-3.3-70B-Instruct at Q4 needs ≈ 46.83 GB even at 4K context, beyond the L40S 48GB's 48 GB GDDR6. It would only run with heavy CPU/disk offload. Decode at Q4/8K: ≈ 17.8 tok/s (estimated, batch 1). FP16 needs ≈ 156.2 GB — does not fit on this card. Estimates come from memory-bandwidth math; rows tagged measured override estimates. Relative ranking is more reliable than absolute tok/s. Methodology.

Fit & speed by quant and context

Quant	Context	VRAM needed	Fits?	tok/s (decode)
Q4	4K	46.83 GB	No	23.6 est.
Q4	8K default	53.83 GB	No	17.8 est.
Q4	32K	95.83 GB	No	5.6 est.
FP16	4K	149.2 GB	No	0.6 est.
FP16	8K default	156.2 GB	No	0.6 est.
FP16	32K	198.2 GB	No	—

Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s means the model is far beyond this card (offload-only territory).

VRAM breakdown at 8K context

Quant	Weights	KV cache	Overhead	Total needed	L40S 48GB VRAM
Q4	38.83 GB	14.0 GB	1 GB	53.83 GB	48 GB
FP16	141.2 GB	14.0 GB	1 GB	156.2 GB	48 GB

Other GPUs that run Llama-3.3-70B-Instruct

• B200 192GB SXM — ≈72.7 tok/s Q4 est.
• MI300X 192GB — ≈48.2 tok/s Q4 est.
• H200 141GB SXM5 — ≈43.6 tok/s Q4 est.
• H100 80GB SXM5 — ≈30.5 tok/s Q4 est.
• Gaudi 3 128GB — ≈29.8 tok/s Q4 est.
• A100 80GB SXM4 — ≈18.5 tok/s Q4 est.
• Mac Studio M3 Ultra 96GB — ≈7.4 tok/s Q4 est.
• Mac Studio M2 Ultra 192GB — ≈7.3 tok/s Q4 est.

All GPUs for Llama-3.3-70B-Instruct →

Other models for the L40S 48GB

• Qwen3-0.6B — ≈687.3 tok/s Q4 est.
• gemma-3-1b-it — ≈549.8 tok/s Q4 est.
• Llama-3.2-1B-Instruct — ≈458.2 tok/s Q4 est.
• Qwen3-1.7B — ≈274.9 tok/s Q4 est.
• Qwen3-30B-A3B — ≈183.3 tok/s Q4 est.
• Qwen3-Coder-30B-A3B-Instruct — ≈183.3 tok/s Q4 est.
• GLM-4.7-Flash — ≈183.3 tok/s Q4 est.
• Qwen3-30B-A3B-Instruct-2507 — ≈183.3 tok/s Q4 est.

All models on the L40S 48GB →