Best GPU for Gemma 3 27B IT
The smallest memory budget in our catalog that runs Gemma 3 27B IT cleanly at 8,192 tokens is 22 GiB net, with Q4_0 (17.2 GiB–18.0 GiB total). Cards at that budget: GeForce RTX 3090 24GB, GeForce RTX 4090 24GB, Radeon RX 7900 XTX 24GB. The largest file, BF16, is comfortable from 96 GiB net: Ryzen AI Max+ 395.
Which card, by net memory
| Net budget | Devices | Best file that runs | Total at 8,192 | Result | Same file at 32,768 |
|---|---|---|---|---|---|
| 18 GiB | Radeon RX 7900 XT 20GB | Q4_0 | 17.2 GiB–18.0 GiB | Tight | No |
| 22 GiB | GeForce RTX 3090 24GB, GeForce RTX 4090 24GB, Radeon RX 7900 XTX 24GB | Q4_0 | 17.2 GiB–18.0 GiB | Fits | Tight |
| 30 GiB | GeForce RTX 5090 32GB | Q4_0 | 17.2 GiB–18.0 GiB | Comfortable | Comfortable |
| 46 GiB | NVIDIA L40S 48GB | Q4_0 | 17.2 GiB–18.0 GiB | Comfortable | Comfortable |
| 54 GiB | Apple M2 Max 64GB unified (MacBook Pro / Mac Studio) | Q4_0 | 17.2 GiB–18.0 GiB | Comfortable | Comfortable |
| 78 GiB | NVIDIA A100 80GB SXM, NVIDIA H100 80GB SXM | BF16 | 52.8 GiB–54.7 GiB | Fits | Fits |
| 96 GiB | Ryzen AI Max+ 395 | BF16 | 52.8 GiB–54.7 GiB | Comfortable | Comfortable |
Rows start at the first budget where the smallest official file is at least tight and stop at the first where the largest file is comfortable. Devices in one row have the same net budget (capacity minus the catalog's default reserve), so they get the same verdict; choose between them on bandwidth, software support and price, which this page does not rank. llama.cpp, one sequence, fp16 KV cache, typical extras. Unified-memory devices use the reserve in their catalog record. "n/a" in the last column means the model's native window is shorter than 32,768 tokens.
Files this ranking uses
| File | Weights | Source repository |
|---|---|---|
| BF16 | 51.10 GiB (54,864,980,440 bytes) | google/gemma-3-27b-it |
| Q4_0 | 16.05 GiB (17,229,630,496 bytes) | google/gemma-3-27b-it-qat-q4_0-gguf |
Byte counts come from the repository file listings recorded on the model page. Memory estimates are formulas, not measurements on these cards. No speed figure is given: none was measured here.
Data sources
- Gemma 3 model card (Google AI for Developers) — Input context 128K tokens for 4B, 12B, 27B and 32K for 1B; images encoded to 256 tokens each. License: Gemma Terms of Use.
- Google gemma_pytorch config.py (architecture) — get_config_for_27b_v3: num_hidden_layers=62, num_attention_heads=32, num_key_value_heads=16, hidden_size=5376, head_dim=128, sliding_window_size=1024, attn_types = 5 LOCAL_SLIDING then 1 GLOBAL, cycled by layer index (gemma/model.py: attn_types[i % len(attn_types)]), so 10 global and 52 sliding layers. The Hugging Face config.json is gated (HTTP 401), so it was not read.
- Cross-check only: llama.cpp org GGUF header converted from Google's config (not an official Google source) — Header read with an HTTP range request on 2026-10-18: block_count 62, head_count 32, head_count_kv 16, embedding_length 5376, key_length 128, sliding_window 1024, context_length 131072. All agree with gemma_pytorch config.py.
- Method and reserves — catalog dataset 2026-10-02.1; hardware capacities from vendor spec pages listed on each device page