| --- |
| tags: |
| - benchmark |
| - amd |
| - rocm |
| - local-llm |
| - linux |
| - latam |
| language: |
| - es |
| - en |
| license: apache-2.0 |
| --- |
| |
| # AMD RX 6700 XT + ROCm — Master Benchmark |
|
|
| > 🇲🇽 [Versión en Español](#versión-en-español) | 🇺🇸 [English Version](#english-version) |
|
|
| --- |
|
|
| ## English Version |
|
|
| ### 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 fix | HSA_OVERRIDE_GFX_VERSION=10.3.0 | |
|
|
| ### Results (average of 3 runs) |
|
|
| | Model | Category | Prefill (tok/s) | Decode (tok/s) | VRAM | |
| |-------|----------|----------------|----------------|------| |
| | llama3.2:3b | General / 3B / Dense | 1165.30 | 107.79 | ✅ full GPU | |
| | gemma4:e2b | General / 2B / Dense | 941.24 | 84.69 | ✅ full GPU | |
| | qwen2.5:7b | General / 7B / Dense | 1003.74 | 61.09 | ✅ full GPU | |
| | qwen2.5-coder:7b | Coding / 7B / Dense | 1073.21 | 60.86 | ✅ full GPU | |
| | mistral:7b | General / 7B / Dense | 431.96 | 60.36 | ✅ full GPU | |
| | gemma4:e4b | General / 4B / Dense | 334.35 | 21.23 | ⚠️ partial | |
| | phi3:mini | General / 3.8B / Dense | 234.12 | 20.24 | ⚠️ partial | |
| | gemma4:26b | General / 26B / MoE | 141.55 | 9.50 | ⚠️ RAM offload | |
| | deepseek-r1:14b | Reasoning / 14B / Dense | 59.48 | 4.91 | ⚠️ RAM offload | |
| | qwen2.5:14b | General / 14B / Dense | 162.33 | 4.84 | ⚠️ RAM offload | |
| | qwopus | Reasoning / 27B / IQ3_XS | 7.09 | 3.76 | ⚠️ RAM offload | |
| |
| > **Key insight:** 7B models that fit in VRAM hit 60+ tok/s consistently. |
| > 14B+ models drop to ~5 tok/s due to RAM offload on 12GB VRAM. |
| |
| ### Which model for which task? |
| |
| #### Fast chat / quick answers |
| **Best:** `llama3.2:3b` (107 tok/s) or `gemma4:e2b` (84 tok/s) |
| - Use when: you need instant responses, simple Q&A, quick summaries |
| - Avoid when: complex reasoning or long code generation needed |
| ```bash |
| ollama run llama3.2:3b "Summarize this in one paragraph: ..." |
| ``` |
| |
| #### General purpose / daily use |
| **Best:** `qwen2.5:7b` (61 tok/s) |
| - Use when: writing, analysis, explanations, multilingual tasks |
| - Context: handles Spanish and English well |
| - Tip: temperature 0.7 for creative tasks, 0.1 for factual |
| ```bash |
| ollama run qwen2.5:7b "Explain the difference between DevOps and MLOps" |
| ``` |
| |
| #### Coding assistant |
| **Best:** `qwen2.5-coder:7b` (60 tok/s) |
| - Use when: code generation, debugging, code review, shell scripts |
| - Tip: always include the full error message, not just the line |
| - Works well with: Python, Bash, Terraform, Docker |
| ```bash |
| ollama run qwen2.5-coder:7b "Write a Python script that monitors GPU usage every 5 seconds" |
| ``` |
| |
| #### Deep reasoning / complex problems |
| **Best:** `deepseek-r1:14b` (4.91 tok/s) — slow but worth it |
| - Use when: math problems, architecture decisions, multi-step analysis |
| - Note: uses `<think>` tags internally, expect longer responses |
| - Tip: be patient, the thinking process is where the quality comes from |
| ```bash |
| ollama run deepseek-r1:14b "Design a fault-tolerant MLOps pipeline for a startup with limited budget" |
| ``` |
| |
| #### Gemma 4 family (Google, April 2026) |
| - `gemma4:e2b`: fastest of the family, good for edge/interactive use |
| - `gemma4:e4b`: better quality, still usable speed on 12GB |
| - `gemma4:26b`: MoE architecture, activates only 4B params per token, but RAM offload kills speed on 12GB VRAM |
| ```bash |
| ollama run gemma4:e4b "What is sliding window attention and why does it matter?" |
| ``` |
| |
| #### Qwopus — Claude Opus reasoning distilled into Qwen3.5 |
| **`qwopus` = Qwen3.5-27B fine-tuned on Claude 4.6 Opus reasoning chains** |
| - Very slow on 12GB (3.76 tok/s) due to heavy RAM offload |
| - Quality is noticeably better than base Qwen for structured reasoning |
| - Best use: offline complex analysis where you can wait |
| - Tip: set num_ctx to 4096 max on 12GB to avoid further slowdown |
| ```bash |
| ollama run qwopus "Analyze the tradeoffs between serverless and container-based MLOps" |
| ``` |
|
|
| ### ROCm optimization tips for RX 6700 XT |
|
|
| ```bash |
| # Always export this before running ollama manually |
| export HSA_OVERRIDE_GFX_VERSION=10.3.0 |
| |
| # In ollama service (add to /etc/systemd/system/ollama.service.d/rocm-fix.conf) |
| Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0" |
| Environment="ROCR_VISIBLE_DEVICES=0" |
| |
| # Context length vs speed tradeoff |
| # 12GB VRAM fills fast with large context |
| # For 7B models: 16K context is fine |
| # For 14B+ models: keep at 4K-8K max or decode drops further |
| |
| # Verify GPU is being used |
| journalctl -u ollama -n 5 --no-pager | grep "AMD Radeon" |
| ``` |
|
|
| ### VRAM sweet spot guide |
|
|
| | Model size | Fits in 12GB? | Expected decode | |
| |-----------|---------------|----------------| |
| | 2-4B Q4 | ✅ Yes | 80-110 tok/s | |
| | 7B Q4 | ✅ Yes | 55-65 tok/s | |
| | 8B Q4 | ✅ Yes | 50-60 tok/s | |
| | 14B Q4 | ❌ RAM offload | 4-6 tok/s | |
| | 27B IQ3 | ❌ RAM offload | 3-5 tok/s | |
| | 26B MoE Q4 | ❌ RAM offload | 8-10 tok/s | |
|
|
| --- |
|
|
| ## Versión en Español |
|
|
| ### 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 | |
| | Fix ROCm | HSA_OVERRIDE_GFX_VERSION=10.3.0 | |
|
|
| ### Resultados (promedio de 3 corridas) |
|
|
| | Modelo | Categoría | Prefill (tok/s) | Decode (tok/s) | VRAM | |
| |--------|-----------|----------------|----------------|------| |
| | llama3.2:3b | General / 3B / Dense | 1165.30 | 107.79 | ✅ GPU completa | |
| | gemma4:e2b | General / 2B / Dense | 941.24 | 84.69 | ✅ GPU completa | |
| | qwen2.5:7b | General / 7B / Dense | 1003.74 | 61.09 | ✅ GPU completa | |
| | qwen2.5-coder:7b | Coding / 7B / Dense | 1073.21 | 60.86 | ✅ GPU completa | |
| | mistral:7b | General / 7B / Dense | 431.96 | 60.36 | ✅ GPU completa | |
| | gemma4:e4b | General / 4B / Dense | 334.35 | 21.23 | ⚠️ parcial | |
| | phi3:mini | General / 3.8B / Dense | 234.12 | 20.24 | ⚠️ parcial | |
| | gemma4:26b | General / 26B / MoE | 141.55 | 9.50 | ⚠️ offload a RAM | |
| | deepseek-r1:14b | Reasoning / 14B / Dense | 59.48 | 4.91 | ⚠️ offload a RAM | |
| | qwen2.5:14b | General / 14B / Dense | 162.33 | 4.84 | ⚠️ offload a RAM | |
| | qwopus | Reasoning / 27B / IQ3_XS | 7.09 | 3.76 | ⚠️ offload a RAM | |
| |
| > **Conclusión clave:** Modelos 7B que caben en VRAM logran 60+ tok/s consistente. |
| > Modelos 14B+ caen a ~5 tok/s por offload a RAM en 12GB VRAM. |
| |
| ### ¿Qué modelo usar para qué? |
| |
| #### Chat rápido / respuestas inmediatas |
| **Mejor:** `llama3.2:3b` (107 tok/s) o `gemma4:e2b` (84 tok/s) |
| - Usar cuando: necesitas respuestas instantáneas, preguntas simples, resúmenes rápidos |
| - Evitar cuando: necesitas razonamiento complejo o código extenso |
| ```bash |
| ollama run llama3.2:3b "Resume esto en un párrafo: ..." |
| ``` |
| |
| #### Uso general / día a día |
| **Mejor:** `qwen2.5:7b` (61 tok/s) |
| - Usar cuando: redacción, análisis, explicaciones, tareas multilingüe |
| - Maneja bien español e inglés |
| - Tip: temperatura 0.7 para tareas creativas, 0.1 para preguntas factuales |
| ```bash |
| ollama run qwen2.5:7b "Explica la diferencia entre DevOps y MLOps" |
| ``` |
| |
| #### Asistente de código |
| **Mejor:** `qwen2.5-coder:7b` (60 tok/s) |
| - Usar cuando: generar código, debuggear, revisar scripts, Terraform, Docker |
| - Tip: siempre incluye el mensaje de error completo, no solo la línea |
| - Funciona bien con: Python, Bash, Terraform, Docker |
| ```bash |
| ollama run qwen2.5-coder:7b "Escribe un script en Python que monitoree el uso de GPU cada 5 segundos" |
| ``` |
| |
| #### Razonamiento profundo / problemas complejos |
| **Mejor:** `deepseek-r1:14b` (4.91 tok/s) — lento pero vale la pena |
| - Usar cuando: matemáticas, decisiones de arquitectura, análisis multi-paso |
| - Nota: usa tags `<think>` internamente, espera respuestas más largas |
| - Tip: ten paciencia, el proceso de pensamiento es donde está la calidad |
| ```bash |
| ollama run deepseek-r1:14b "Diseña un pipeline de MLOps tolerante a fallos para una startup con presupuesto limitado" |
| ``` |
| |
| #### Familia Gemma 4 (Google, abril 2026) |
| - `gemma4:e2b`: el más rápido de la familia, bueno para uso interactivo |
| - `gemma4:e4b`: mejor calidad, velocidad usable en 12GB |
| - `gemma4:26b`: arquitectura MoE, activa solo 4B params por token, pero el offload a RAM mata la velocidad en 12GB |
| ```bash |
| ollama run gemma4:e4b "¿Qué es el sliding window attention y por qué importa?" |
| ``` |
| |
| #### Qwopus — razonamiento de Claude Opus destilado en Qwen3.5 |
| **`qwopus` = Qwen3.5-27B fine-tuned con chains de razonamiento de Claude 4.6 Opus** |
| - Muy lento en 12GB (3.76 tok/s) por offload pesado a RAM |
| - La calidad de razonamiento estructurado es notablemente mejor que el Qwen base |
| - Mejor uso: análisis complejos offline donde puedes esperar |
| - Tip: limita num_ctx a 4096 máximo en 12GB para evitar más lentitud |
| ```bash |
| ollama run qwopus "Analiza los tradeoffs entre serverless y contenedores para MLOps" |
| ``` |
|
|
| ### Tips de optimización ROCm para RX 6700 XT |
|
|
| ```bash |
| # Siempre exporta esto antes de usar ollama manualmente |
| export HSA_OVERRIDE_GFX_VERSION=10.3.0 |
| |
| # En el servicio ollama (agrega a rocm-fix.conf) |
| Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0" |
| Environment="ROCR_VISIBLE_DEVICES=0" |
| |
| # Tradeoff context length vs velocidad |
| # Para modelos 7B: 16K context está bien |
| # Para modelos 14B+: máximo 4K-8K o el decode baja aún más |
| |
| # Verificar que la GPU está siendo usada |
| journalctl -u ollama -n 5 --no-pager | grep "AMD Radeon" |
| ``` |
|
|
| ### Guía de VRAM para 12GB |
|
|
| | Tamaño modelo | ¿Cabe en 12GB? | Decode esperado | |
| |--------------|----------------|----------------| |
| | 2-4B Q4 | ✅ Sí | 80-110 tok/s | |
| | 7B Q4 | ✅ Sí | 55-65 tok/s | |
| | 8B Q4 | ✅ Sí | 50-60 tok/s | |
| | 14B Q4 | ❌ Offload RAM | 4-6 tok/s | |
| | 27B IQ3 | ❌ Offload RAM | 3-5 tok/s | |
| | 26B MoE Q4 | ❌ Offload RAM | 8-10 tok/s | |
|
|
| --- |
| *By [Positronica Labs](https://github.com/G10hdz) — CDMX, México* |
|
|
| [](https://github.com/G10hdz) |
| [](https://www.linkedin.com/in/mayte-giovanna-hernandez-rios/) |
|
|
| *Hardware de segunda mano. Software libre. AI para todos.* |
|
|