AMD · 192 GB HBM3 · 5300 GB/s bandwidth ·
Google · 30.7B · model ctx up to 256K
Yes — the MI300X 192GB (192 GB HBM3) can run gemma-4-31B-it. At Q4 with the default 8K context it needs ≈ 24.03 GB VRAM (weights 16.89 GB + KV cache 6.14 GB + 1 GB runtime overhead), which fits within 192 GB. Decode at Q4/8K: ≈ 109.9 tok/s (estimated, batch 1). FP16 (≈ 68.54 GB) also fits, at ≈ 30.2 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
20.96 GB
Yes
109.9est.
Q4
8K default
24.03 GB
Yes
109.9est.
Q4
32K
42.45 GB
Yes
109.9est.
FP16
4K
65.47 GB
Yes
30.2est.
FP16
8K default
68.54 GB
Yes
30.2est.
FP16
32K
86.96 GB
Yes
30.2est.
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).