what/llm/can/i/run
gpus/nvidia/l40s

L40S

Released 2023-08-08 · Street $8,500
vram
48 GB
discrete
bandwidth
864
GB/s
models fit
17
top · Q8_0
power
350 W
high

▸Spec sheet · top pick

Hardware spec
L40S
NVIDIA · Ada Lovelace · sm_89
vram
48 GB
bandwidth
864 GB/s
tdp
350 W
released
2023-08-08
msrp
$8,000
street
~$8,500
★ Sweet spot for this card
gemma 3 27b
Q8_0 · GGUF · 28.9 GB
bench50.6 · conf A
vram28.9 / 48 GB
speed30 tok/s
Best benchmark per GB of VRAM in your tier. Q8_0wastes <3% quality vs FP16, leaves headroom for 32k context.
Full model inspect →

▸Top 10 models on this GPU

#modelquantvramtok/sbench ↓fit
01gemma 3 27bQ8_028.9 GB3050.6Great
02Qwen2.5 Coder 32BQ8_034.1 GB2548.8Great
03Gemma 2 27BQ8_028.9 GB3046.2Great
04Qwen3 32BQ8_034.1 GB2540.0Great
05Gemma 2 9BFP1619.4 GB4535.0Great
06Gemma 3 12BFP1625.7 GB3431.3Great
07deepseek r1 distill llama 70bQ5_K_M46.4 GB1951.7Tight
08deepseek r1 distill qwen 32bQ8_034.1 GB2533.3Great
09Gemma 3 4BFP168.9 GB9715.1Great
10Mixtral 8x7B v0.1Q6_K37.3 GB2313.3Great
Showing 10 of 17 models that fit · see full ranked list →

▸Upgrade options

★ Best value
+H100 PCIe
80 GB · ~$28,000 · +$19,500 over L40S
Unlocks +1 models not in current rank
See H100 PCIe →
Bigger tier
+H100 SXM5
80 GB · ~$40,000 · +$31,500 over L40S
Unlocks +1 models not in current rank
See H100 SXM5 →
Premium tier
+H200 SXM5
141 GB · ~$45,000 · +$36,500 over L40S
Unlocks +2 models not in current rank
See H200 SXM5 →

▸Also compare

RTX 5090
32 GB · NVIDIA · 72% cheaper
A6000 Pro
48 GB · NVIDIA · 51% cheaper
RTX 6000 Ada
48 GB · NVIDIA · 18% cheaper
A100 40GB
40 GB · NVIDIA · 6% cheaper
Not running a L40S?
Tell us what you actually have — we'll re-rank everything for your machine.
Change my machine