| --- |
| tags: |
| - benchmark |
| - amd |
| - rocm |
| - gemma4 |
| - local-llm |
| - linux |
| - spanish |
| - latam |
| language: |
| - es |
| - en |
| license: apache-2.0 |
| --- |
| |
| # Gemma 4 on AMD RX 6700 XT + ROCm |
|
|
| > 🇲🇽 [Versión en Español](#versión-en-español) | 🇺🇸 [English Version](#english-version) |
|
|
| Benchmarks of Google's Gemma 4 (April 2026) running on AMD GPU with ROCm on Linux. |
| **This documentation doesn't exist in Spanish — for the LatAm community.** |
|
|
| --- |
|
|
| ## English Version |
|
|
| ### Why this matters |
|
|
| Most local AI guides assume: |
| - NVIDIA GPU (CUDA) |
| - Apple Silicon (macOS) |
| - $1,500+ USD budget |
|
|
| This repo documents that a **budget AMD GPU on Linux** running Gemma 4 — released April 2026 — outperforms a brand new Mac Mini M4 in decode speed, at a fraction of the cost. |
|
|
| ### Hardware |
|
|
| | Component | Detail | |
| |-----------|--------| |
| | GPU | AMD RX 6700 XT 12GB (gfx1031) | |
| | CPU | AMD Ryzen 5 5600G | |
| | RAM | 16GB | |
| | OS | Pop!_OS 24.04 LTS | |
| | Ollama | 0.20.2 | |
| | ROCm override | HSA_OVERRIDE_GFX_VERSION=10.3.0 | |
|
|
| ### Results (average of 3 runs) |
|
|
| | Model | Prefill (tok/s) | Decode (tok/s) | VRAM | Status | |
| |-------|----------------|----------------|------|--------| |
| | gemma4:e2b | 1022.39 | 85.00 | ~2GB | ROCm ✅ full GPU | |
| | gemma4:e4b | 697.91 | 57.34 | ~4GB | ROCm ✅ full GPU | |
| | gemma4:26b | 157.94 | 10.70 | 12GB + RAM offload | ROCm ⚠️ partial | |
|
|
| > **Note:** Prefill varies between runs due to KV cache warming. |
| > Decode is the number that matters for user experience — very consistent. |
|
|
| ### Platform comparison (decode ~4B models) |
|
|
| | Hardware | Decode tok/s | Price (approx USD) | |
| |----------|--------------|--------------------| |
| | **RX 6700 XT + ROCm (this repo)** | **57-85** | existing hardware | |
| | Mac Mini M4 16GB | 25-40 | ~$600 new | |
| | RTX 3070 12GB | 50-80 | ~$500 used | |
|
|
| **The RX 6700 XT doubles the Mac Mini M4 16GB in decode speed for models that fit in VRAM.** |
|
|
| ### The gfx1031 problem |
|
|
| The RX 6700 XT reports `gfx1031` architecture, but ROCm doesn't officially support it. |
| Without the override, Ollama falls back to CPU (~3-5 tok/s). |
|
|
| **Fix:** |
| ```bash |
| sudo mkdir -p /etc/systemd/system/ollama.service.d |
| |
| sudo tee /etc/systemd/system/ollama.service.d/rocm-fix.conf << 'CONF' |
| [Service] |
| Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0" |
| Environment="ROCR_VISIBLE_DEVICES=0" |
| CONF |
| |
| sudo systemctl daemon-reload |
| sudo systemctl restart ollama |
| ``` |
|
|
| **Verify it worked:** |
| ```bash |
| journalctl -u ollama -n 10 --no-pager | grep "AMD Radeon" |
| # Expected output: |
| # description="AMD Radeon RX 6700 XT" total="12.0 GiB" available="11.1 GiB" |
| ``` |
|
|
| ### Full setup |
|
|
| ```bash |
| # 1. Install Ollama |
| curl -fsSL https://ollama.com/install.sh | sh |
| |
| # 2. Apply ROCm fix (see above) |
| |
| # 3. Pull models |
| ollama pull gemma4:e2b # 2B - fast, fits easily |
| ollama pull gemma4:e4b # 4B - great quality, fits in VRAM |
| ollama pull gemma4:26b # 26B - requires RAM offload on 12GB |
| |
| # 4. Run |
| ollama run gemma4:e4b "Hello, what is MLOps?" |
| ``` |
|
|
| ### Run the benchmark yourself |
|
|
| ```bash |
| chmod +x bench_gemma4.sh |
| ./bench_gemma4.sh |
| ``` |
|
|
| Runs 3 iterations per model, saves results to `gemma4_benchmark_results.md`. |
|
|
| --- |
|
|
| ## Versión en Español |
|
|
| ### Por qué esto importa |
|
|
| La mayoría de guías de AI local asumen: |
| - GPU NVIDIA (CUDA) |
| - Apple Silicon (macOS) |
| - Presupuesto de $1,500 USD o más |
|
|
| Este repo documenta que una **GPU AMD de segunda mano en Linux** corriendo Gemma 4 (lanzada en abril 2026) supera en velocidad a un Mac Mini M4 nuevo, a una fracción del costo. |
|
|
| ### Hardware |
|
|
| | Componente | Detalle | |
| |------------|---------| |
| | GPU | AMD RX 6700 XT 12GB (gfx1031) | |
| | CPU | AMD Ryzen 5 5600G | |
| | RAM | 16GB | |
| | OS | Pop!_OS 24.04 LTS | |
| | Ollama | 0.20.2 | |
| | Override ROCm | HSA_OVERRIDE_GFX_VERSION=10.3.0 | |
|
|
| ### Resultados (promedio de 3 corridas) |
|
|
| | Modelo | Prefill (tok/s) | Decode (tok/s) | VRAM | Estado | |
| |--------|----------------|----------------|------|--------| |
| | gemma4:e2b | 1022.39 | 85.00 | ~2GB | ROCm ✅ GPU completa | |
| | gemma4:e4b | 697.91 | 57.34 | ~4GB | ROCm ✅ GPU completa | |
| | gemma4:26b | 157.94 | 10.70 | 12GB + offload a RAM | ROCm ⚠️ parcial | |
|
|
| > **Nota:** El prefill varía entre corridas por el KV cache warming. |
| > El decode es el número relevante para experiencia de usuario y es muy consistente. |
|
|
| ### Comparativa vs otras plataformas (decode modelos ~4B) |
|
|
| | Hardware | Decode tok/s | Precio aprox MXN | |
| |----------|-------------|-----------------| |
| | **RX 6700 XT + ROCm (este repo)** | **57-85** | hardware existente | |
| | Mac Mini M4 16GB | 25-40 | ~$23,000 MXN nuevo | |
| | RTX 3070 12GB | 50-80 | ~$10,000 MXN usado | |
|
|
| **La RX 6700 XT dobla en velocidad al Mac Mini M4 16GB en modelos que caben en VRAM.** |
|
|
| ### El problema del gfx1031 |
|
|
| La RX 6700 XT reporta arquitectura `gfx1031` pero ROCm no la tiene en su lista de soporte oficial. |
| Sin el override, Ollama cae a CPU (~3-5 tok/s). |
|
|
| **Fix:** |
| ```bash |
| sudo mkdir -p /etc/systemd/system/ollama.service.d |
| |
| sudo tee /etc/systemd/system/ollama.service.d/rocm-fix.conf << 'CONF' |
| [Service] |
| Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0" |
| Environment="ROCR_VISIBLE_DEVICES=0" |
| CONF |
| |
| sudo systemctl daemon-reload |
| sudo systemctl restart ollama |
| ``` |
|
|
| **Verificar que funcionó:** |
| ```bash |
| journalctl -u ollama -n 10 --no-pager | grep "AMD Radeon" |
| # Output esperado: |
| # description="AMD Radeon RX 6700 XT" total="12.0 GiB" available="11.1 GiB" |
| ``` |
|
|
| ### Setup completo |
|
|
| ```bash |
| # 1. Instalar Ollama |
| curl -fsSL https://ollama.com/install.sh | sh |
| |
| # 2. Aplicar fix de ROCm (ver arriba) |
| |
| # 3. Descargar modelos |
| ollama pull gemma4:e2b # 2B - rápido, cabe fácil |
| ollama pull gemma4:e4b # 4B - buena calidad, cabe en VRAM |
| ollama pull gemma4:26b # 26B - requiere offload a RAM en 12GB |
| |
| # 4. Correr |
| ollama run gemma4:e4b "Hola, explica qué es MLOps" |
| ``` |
|
|
| ### Correr el benchmark tú mismo |
|
|
| ```bash |
| chmod +x bench_gemma4.sh |
| ./bench_gemma4.sh |
| ``` |
|
|
| Corre 3 iteraciones por modelo y guarda los resultados en `gemma4_benchmark_results.md`. |
|
|
| ### Modelos probados |
|
|
| | Modelo | Parámetros | Tipo | ¿Cabe en 12GB? | |
| |--------|-----------|------|----------------| |
| | gemma4:e2b | 2B | Dense | ✅ GPU completa | |
| | gemma4:e4b | 4B | Dense | ✅ GPU completa | |
| | gemma4:26b | 26B (4B activos, MoE) | MoE | ⚠️ offload parcial | |
|
|
| --- |
|
|
| ## About / Sobre este repo |
|
|
| Made in CDMX with second-hand hardware and free software. |
| Hecho en CDMX con hardware de segunda mano y software libre. |
|
|
| *[Positronica Labs](https://github.com/G10hdz) — Building AI tools for Latin America.* |
| *Hardware de segunda mano. Software libre. AI para todos.* |
|
|
