NVIDIA · 48 GB GDDR6 · 864 GB/s bandwidth ·
Google · 11.95B · model ctx up to 256K
Yes — the L40S 48GB (48 GB GDDR6) can run gemma-4-12b-it. At Q4 with the default 8K context it needs ≈ 9.96 GB VRAM (weights 6.57 GB + KV cache 2.39 GB + 1 GB runtime overhead), which fits within 48 GB. Decode at Q4/8K: ≈ 46.0 tok/s (estimated, batch 1). FP16 (≈ 27.29 GB) also fits, at ≈ 12.7 tok/s (estimated). 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
8.77 GB
Yes
46.0est.
Q4
8K default
9.96 GB
Yes
46.0est.
Q4
32K
17.13 GB
Yes
46.0est.
FP16
4K
26.09 GB
Yes
12.7est.
FP16
8K default
27.29 GB
Yes
12.7est.
FP16
32K
34.46 GB
Yes
12.7est.
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).