Gemma 4 (E2B and E4B) - M5 Max 36GB (binned) and Ubuntu 3800x 3060ti llama.cpp benchmark

llama_bench_gemma.sh

# Apple Silicon llama-bench: Gemma 4 (E2B and E4B) Q4_K_M depth sweep
# Testing for Gemma 4 (E2B and E4B) on Apple Silicon

llama-bench \
  -m ~/.lmstudio/models/lmstudio-community/gemma-4-E2B-it-GGUF/gemma-4-E2B-it-Q4_K_M.gguf  \
  -m ~/.lmstudio/models/lmstudio-community/gemma-4-E4B-it-GGUF/gemma-4-E4B-it-Q4_K_M.gguf      \
  -p 512 -n 128 -fa 1 -ngl 99 \
  -d 0,4096,8192,16384 \
  -o md > local-gemma-q4km-m5_max_36gb.md 2>/dev/null

Ubuntu 3800x 32GB 3060ti

model	size	params	backend	ngl	fa	test	t/s
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	pp512	6044.38 ± 382.66
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	tg128	159.16 ± 1.67
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	pp512 @ d4096	5289.56 ± 100.77
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	tg128 @ d4096	154.08 ± 0.24
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	pp512 @ d8192	4698.99 ± 75.61
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	tg128 @ d8192	150.27 ± 0.16
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	pp512 @ d16384	3849.82 ± 52.89
gemma4 E2B Q4_K - Medium	3.18 GiB	4.65 B	CUDA	99	1	tg128 @ d16384	144.19 ± 0.25
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	pp512	3538.31 ± 150.44
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	tg128	94.83 ± 0.10
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	pp512 @ d4096	3119.65 ± 56.53
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	tg128 @ d4096	90.50 ± 0.52
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	pp512 @ d8192	2836.10 ± 35.33
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	tg128 @ d8192	88.40 ± 0.47
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	pp512 @ d16384	2460.92 ± 29.41
gemma4 E4B Q4_K - Medium	4.95 GiB	7.52 B	CUDA	99	1	tg128 @ d16384	84.19 ± 0.34