Intel · 128 GB HBM2e · 3276 GB/s bandwidth ·
Google · 5.1B · model ctx up to 128K
Yes — the Gaudi 3 128GB (128 GB HBM2e) can run gemma-4-E2B-it. At Q4 with the default 8K context it needs ≈ 4.83 GB VRAM (weights 2.81 GB + KV cache 1.02 GB + 1 GB runtime overhead), which fits within 128 GB. Decode at Q4/8K: ≈ 408.8 tok/s (estimated, batch 1). FP16 (≈ 12.22 GB) also fits, at ≈ 112.4 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
4.32 GB
Yes
408.8est.
Q4
8K default
4.83 GB
Yes
408.8est.
Q4
32K
7.88 GB
Yes
408.8est.
FP16
4K
11.71 GB
Yes
112.4est.
FP16
8K default
12.22 GB
Yes
112.4est.
FP16
32K
15.28 GB
Yes
112.4est.
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).