Gemma · Gemma 3 · 27.4B parameters

Gemma 3 27B VRAM requirements

Gemma 3 27B has 62 layers and uses grouped-query attention (16 KV heads). At Q4_K_M the weights come to 15.6 GB, and the best quantisation that fits a 24 GB card is Q5_K_M.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

17.5 GB of 21.8 GB · 80%

022 GB

Weights 15.6 GB

KV cache 1.1 GB

Runtime overhead 0.8 GB

Gemma 3 27B at Q4_K_M leaves 4.3 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation36.5tok/s

Prompt processing1265tok/s

Max context57Ktokens

KV per 1K tokens0GB

Every quantisation of Gemma 3 27B on a RTX 4090

Highlighted row is the highest quality that still fits at 8K context.

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	51.0 GB	53.0 GB	Won't fit	—	1.16
INT8 / W8A8	8.50	26.8 GB	28.7 GB	Won't fit	—	4.33
Q8_0 (GGUF)	8.50	26.8 GB	28.7 GB	Won't fit	—	4.33
FP8 (E4M3)	8.00	25.5 GB	27.4 GB	Won't fit	—	5.29
Q6_K	6.56	20.9 GB	22.9 GB	Won't fit	—	14.2
Q5_K_M	5.67	18.1 GB	20.0 GB	Fits, but tight	28K	31.7
Q5_K_S	5.52	17.6 GB	19.5 GB	Runs comfortably	34K	32.5
Q4_K_M	4.85	15.6 GB	17.5 GB	Runs comfortably	57K	36.5
AWQ 4-bit	4.25	15.5 GB	17.4 GB	Runs comfortably	59K	36.7
GPTQ 4-bit	4.25	15.5 GB	17.4 GB	Runs comfortably	59K	36.7
MXFP4	4.25	15.5 GB	17.4 GB	Runs comfortably	59K	36.7
Q4_K_S	4.58	14.8 GB	16.7 GB	Runs comfortably	67K	38.4
Q4_0	4.55	14.7 GB	16.6 GB	Runs comfortably	68K	38.7
IQ4_XS	4.25	13.8 GB	15.7 GB	Runs comfortably	79K	41.0
Q3_K_M	3.91	12.7 GB	14.7 GB	Runs comfortably	91K	44.1
IQ3_M	3.70	12.1 GB	14.0 GB	Runs comfortably	98K	46.3
IQ3_XXS	3.06	10.2 GB	12.1 GB	Runs comfortably	121K	54.3
Q2_K	2.63	8.9 GB	10.8 GB	Runs comfortably	128K	61.5
IQ2_XXS	2.06	7.1 GB	9.1 GB	Runs comfortably	128K	74.6
IQ1_M	1.75	6.2 GB	8.1 GB	Runs comfortably	128K	84.4

Gemma 3 27B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	136
A100 80GB	80	2039	Runs comfortably	128K	76.0
RTX 5090	32	1792	Runs comfortably	128K	70.7
RTX 4090	24	1008	Runs comfortably	57K	36.5
RTX 3090	24	936	Runs comfortably	57K	35.4
Radeon RX 7900 XTX	24	960	Runs comfortably	57K	33.1
L40S	48	864	Runs comfortably	128K	31.4
RTX A6000	48	768	Runs comfortably	128K	29.1
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	28.7
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	21.4
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	11.0
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	10.7
RTX 5080	16	960	Won't fit	—	9.29
RTX 5070 Ti	16	896	Won't fit	—	9.17
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	8.21
RTX 4080 Super	16	736	Won't fit	—	7.98
RTX 4070 Ti Super	16	672	Won't fit	—	7.81
RTX 5060 Ti 16GB	16	448	Won't fit	—	7.67
RTX 4060 Ti 16GB	16	288	Won't fit	—	5.89
RTX 5070	12	672	Won't fit	—	5.12
RTX 4070 Super	12	504	Won't fit	—	4.49
RTX 4070	12	504	Won't fit	—	4.49
RTX 3060 12GB	12	360	Won't fit	—	4.45
RTX 3080 10GB	10	760	Won't fit	—	3.97
Arc B580	12	456	Won't fit	—	3.75

Architecture

Parameters	27.4B
Layers	62
Hidden size	5376
Attention heads / KV heads	32 / 16
Head dimension	128
Vocabulary	262,144
Trained context	128K
Sliding window	1024 (every 6th layer is global)
KV cache per 1K tokens	0 GB
Hugging Face	google/gemma-3-27b-it

The Gemma family

Google's open models. Two things dominate the memory: a 262k vocabulary — on Gemma 3 1B the embedding table is about a third of the file — and sliding-window attention, where only every sixth layer of Gemma 3 sees the full context, and every second layer on Gemma 2. The cache grows far more slowly than the context length suggests.

huggingface.co/google · ai.google.dev/gemma · all 6 Gemma models

Gemma 3 1B
1B · 26 layers Gemma 3 4B
4.3B · 34 layers Gemma 3 12B
12.2B · 48 layers Gemma 2 9B
9.2B · 42 layers Gemma 2 27B
27.2B · 46 layers

Direct answers

Gemma 3 27B on RTX 5090
See the verdict Gemma 3 27B on RTX 4090
See the verdict Gemma 3 27B on RTX 3090
See the verdict Gemma 3 27B on RTX 5080
See the verdict Gemma 3 27B on RTX 5070 Ti
See the verdict Gemma 3 27B on RTX 5070
See the verdict Gemma 3 27B on RTX 4070 Ti Super
See the verdict Gemma 3 27B on RTX 4070 Super
See the verdict Gemma 3 27B on RTX 4070
See the verdict