Gemma · Gemma 2 · 9.2B parameters
Gemma 2 9B VRAM requirements
Gemma 2 9B has 42 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 5.3 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.
Runs comfortably
8.1 GB of 21.8 GB · 37%Gemma 2 9B at Q4_K_M leaves 13.7 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of Gemma 2 9B 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 | 17.2 GB | 20.0 GB | Fits, but tight | 8K | 32.4 |
| INT8 / W8A8 | 8.50 | 8.9 GB | 11.7 GB | Runs comfortably | 8K | 58.1 |
| Q8_0 (GGUF) | 8.50 | 8.9 GB | 11.7 GB | Runs comfortably | 8K | 58.1 |
| FP8 (E4M3) | 8.00 | 8.6 GB | 11.4 GB | Runs comfortably | 8K | 60.0 |
| Q6_K | 6.56 | 7.1 GB | 9.8 GB | Runs comfortably | 8K | 70.9 |
| Q5_K_M | 5.67 | 6.1 GB | 8.9 GB | Runs comfortably | 8K | 79.9 |
| Q5_K_S | 5.52 | 5.9 GB | 8.7 GB | Runs comfortably | 8K | 81.7 |
| AWQ 4-bit | 4.25 | 5.8 GB | 8.6 GB | Runs comfortably | 8K | 82.9 |
| GPTQ 4-bit | 4.25 | 5.8 GB | 8.6 GB | Runs comfortably | 8K | 82.9 |
| MXFP4 | 4.25 | 5.8 GB | 8.6 GB | Runs comfortably | 8K | 82.9 |
| Q4_K_M | 4.85 | 5.3 GB | 8.1 GB | Runs comfortably | 8K | 89.5 |
| Q4_K_S | 4.58 | 5.0 GB | 7.8 GB | Runs comfortably | 8K | 93.1 |
| Q4_0 | 4.55 | 5.0 GB | 7.8 GB | Runs comfortably | 8K | 93.6 |
| IQ4_XS | 4.25 | 4.7 GB | 7.5 GB | Runs comfortably | 8K | 98.0 |
| Q3_K_M | 3.91 | 4.4 GB | 7.2 GB | Runs comfortably | 8K | 103 |
| IQ3_M | 3.70 | 4.2 GB | 7.0 GB | Runs comfortably | 8K | 107 |
| IQ3_XXS | 3.06 | 3.6 GB | 6.3 GB | Runs comfortably | 8K | 120 |
| Q2_K | 2.63 | 3.1 GB | 5.9 GB | Runs comfortably | 8K | 131 |
| IQ2_XXS | 2.06 | 2.6 GB | 5.4 GB | Runs comfortably | 8K | 149 |
| IQ1_M | 1.75 | 2.3 GB | 5.1 GB | Runs comfortably | 8K | 161 |
Gemma 2 9B 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 | 8K | 326 |
| A100 80GB | 80 | 2039 | Runs comfortably | 8K | 184 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 8K | 172 |
| RTX 5080 | 16 | 960 | Runs comfortably | 8K | 93.9 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 8K | 89.5 |
| RTX 5070 Ti | 16 | 896 | Runs comfortably | 8K | 87.8 |
| RTX 3090 | 24 | 936 | Runs comfortably | 8K | 86.8 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 8K | 81.1 |
| L40S | 48 | 864 | Runs comfortably | 8K | 77.0 |
| RTX A6000 | 48 | 768 | Runs comfortably | 8K | 71.5 |
| RTX 3080 10GB | 10 | 760 | Fits, but tight | 8K | 70.8 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 8K | 70.4 |
| RTX 5070 | 12 | 672 | Runs comfortably | 8K | 66.2 |
| RTX 4080 Super | 16 | 736 | Runs comfortably | 8K | 65.8 |
| RTX 4070 Ti Super | 16 | 672 | Runs comfortably | 8K | 60.1 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 8K | 52.5 |
| RTX 4070 Super | 12 | 504 | Runs comfortably | 8K | 45.3 |
| RTX 4070 | 12 | 504 | Runs comfortably | 8K | 45.3 |
| RTX 5060 Ti 16GB | 16 | 448 | Runs comfortably | 8K | 44.4 |
| Arc B580 | 12 | 456 | Runs comfortably | 8K | 34.8 |
| RTX 3060 12GB | 12 | 360 | Runs comfortably | 8K | 33.9 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 8K | 27.2 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 8K | 26.4 |
| RTX 4060 Ti 16GB | 16 | 288 | Runs comfortably | 8K | 26.0 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 8K | 20.3 |
Architecture
| Parameters | 9.2B |
| Layers | 42 |
| Hidden size | 3584 |
| Attention heads / KV heads | 16 / 8 |
| Head dimension | 256 |
| Vocabulary | 256,000 |
| Trained context | 8K |
| Sliding window | 4096 (every 2th layer is global) |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | google/gemma-2-9b-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