Intel · 12 GB GDDR6 · 456 GB/s bandwidth ·
Meta · 1.2B · model ctx up to 128K
Yes — the Arc B580 12GB (12 GB GDDR6) can run Llama-3.2-1B-Instruct. At Q4 with the default 8K context it needs ≈ 1.9 GB VRAM (weights 0.66 GB + KV cache 0.24 GB + 1 GB runtime overhead), which fits within 12 GB. Decode at Q4/8K: ≈ 241.8 tok/s (estimated, batch 1). FP16 (≈ 3.64 GB) also fits, at ≈ 66.5 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
1.78 GB
Yes
241.8est.
Q4
8K default
1.9 GB
Yes
241.8est.
Q4
32K
2.62 GB
Yes
241.8est.
FP16
4K
3.52 GB
Yes
66.5est.
FP16
8K default
3.64 GB
Yes
66.5est.
FP16
32K
4.36 GB
Yes
66.5est.
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).