AMD Radeon RX 9070 GRE
GPUAMD RDNA 4 (a cut-down Navi 48): 12 GB GDDR6 on a 192-bit bus, 432 GB/s of memory bandwidth, 220 W board power. Launched in China on 2025-05-08 and worldwide on 2026-06-02. Its bandwidth is roughly a third below the RX 9070's, which is the binding constraint.
14
Run records
10
Models covered
4
Frameworks covered
L2 Cross-framework verified × 14
Evidence mix
Typical measured performance
Published benchmark averages per model × framework × quantization on this hardware.
- vLLM · IQ1_S Decode 165.6 · Prefill 729.98 ·
- Strata · IQ1_S Decode 212.06 · Prefill 968.79 ·
- llama.cpp · IQ2_M Decode 264.62 · Prefill 925.29 ·
- Strata · IQ1_S Decode 369.89 · Prefill 1215.87 ·
- Ollama · Q2_K Decode 89.93 · Prefill 344.74 ·
- Strata · Q3_K_S Decode 124.72 · Prefill 470.9 ·
- llama.cpp · Q2_K Decode 124.6 · Prefill 437.09 ·
- Ollama · Q8_0 Decode 79.82 · Prefill 368.94 ·
- llama.cpp · Q3_K_M Decode 100.99 · Prefill 537.73 ·
- llama.cpp · Q5_K_S Decode 85.62 · Prefill 376.76 ·
- Ollama · IQ3_XS Decode 244.89 · Prefill 1100.19 ·
- Strata · Q5_K_M Decode 128.98 · Prefill 579.29 ·
- llama.cpp · IQ3_S Decode 241.94 · Prefill 776.44 ·
- llama.cpp · Q2_K Decode 298.41 · Prefill 1440.77 ·
Fit check
Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed
Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers
Key specs
- Vendor
- AMD
- Architecture
- RDNA 4
- VRAM
- 12 GB
- Memory bandwidth
- 432 GB/s
- TDP
- 220 W
- Release date
- 2025-05-08
Run records
14 configurations
| Model | Framework | Quantization | Decode | Gen time | Samples |
|---|---|---|---|---|---|
| Qwen3.5-4B | vLLM | IQ1_S | 165.6 tok/s | — | 1 |
| Qwen3.5-4B | Strata | IQ1_S | 212.1 tok/s | — | 1 |
| Qwen3.5-0.8B | llama.cpp | IQ2_M | 264.6 tok/s | — | 1 |
| Qwen3.5-0.8B | Strata | IQ1_S | 369.9 tok/s | — | 1 |
| Qwen3.8-9B | Ollama | Q2_K | 89.9 tok/s | — | 1 |
| Qwen3-8B | Strata | Q3_K_S | 124.7 tok/s | — | 1 |
| Qwen3-4B | llama.cpp | Q2_K | 124.6 tok/s | — | 1 |
| Qwen3-1.7B | Ollama | Q8_0 | 79.8 tok/s | — | 1 |
| Qwen3-4B-Thinking-2507 | llama.cpp | Q3_K_M | 101 tok/s | — | 1 |
| Qwen3-4B-Thinking-2507 | llama.cpp | Q5_K_S | 85.6 tok/s | — | 1 |
| Qwen2.5-0.5B-Instruct | Ollama | IQ3_XS | 244.9 tok/s | — | 1 |
| Qwen2.5-Omni-3B | Strata | Q5_K_M | 129 tok/s | — | 1 |
| Qwen3-Embedding-0.6B | llama.cpp | IQ3_S | 241.9 tok/s | — | 1 |
| Qwen3-Embedding-0.6B | llama.cpp | Q2_K | 298.4 tok/s | — | 1 |