NVIDIA · 192 GB HBM3e · 8000 GB/s bandwidth ·
Microsoft · 14.7B · model ctx up to 32K
Yes — the B200 192GB SXM (192 GB HBM3e) can run Phi-4-reasoning. At Q4 with the default 8K context it needs ≈ 12.03 GB VRAM (weights 8.09 GB + KV cache 2.94 GB + 1 GB runtime overhead), which fits within 192 GB. Decode at Q4/8K: ≈ 346.3 tok/s (estimated, batch 1). FP16 (≈ 33.34 GB) also fits, at ≈ 95.2 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
10.56 GB
Yes
346.3est.
Q4
8K default
12.03 GB
Yes
346.3est.
Q4
32K
20.84 GB
Yes
346.3est.
FP16
4K
31.87 GB
Yes
95.2est.
FP16
8K default
33.34 GB
Yes
95.2est.
FP16
32K
42.16 GB
Yes
95.2est.
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).