gio5464 commited on
Commit
97eb35b
·
verified ·
1 Parent(s): e9f0368

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +10 -215
README.md CHANGED
@@ -3,234 +3,29 @@ tags:
3
  - benchmark
4
  - amd
5
  - rocm
6
- - gemma4
7
  - local-llm
8
  - linux
9
- - spanish
10
- - latam
11
  language:
12
  - es
13
  - en
14
  license: apache-2.0
15
  ---
16
 
17
- # Gemma 4 on AMD RX 6700 XT + ROCm
18
 
19
- > 🇲🇽 [Versión en Español](#versión-en-español) | 🇺🇸 [English Version](#english-version)
 
20
 
21
- Benchmarks of Google's Gemma 4 (April 2026) running on AMD GPU with ROCm on Linux.
22
- **This documentation doesn't exist in Spanish — for the LatAm community.**
23
 
24
- ---
25
-
26
- ## English Version
27
-
28
- ### Why this matters
29
-
30
- Most local AI guides assume:
31
- - NVIDIA GPU (CUDA)
32
- - Apple Silicon (macOS)
33
- - $1,500+ USD budget
34
-
35
- 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.
36
-
37
- ### Hardware
38
-
39
- | Component | Detail |
40
- |-----------|--------|
41
- | GPU | AMD RX 6700 XT 12GB (gfx1031) |
42
- | CPU | AMD Ryzen 5 5600G |
43
- | RAM | 16GB |
44
- | OS | Pop!_OS 24.04 LTS |
45
- | Ollama | 0.20.2 |
46
- | ROCm override | HSA_OVERRIDE_GFX_VERSION=10.3.0 |
47
-
48
- ### Results (average of 3 runs)
49
-
50
- | Model | Prefill (tok/s) | Decode (tok/s) | VRAM | Status |
51
- |-------|----------------|----------------|------|--------|
52
- | gemma4:e2b | 1022.39 | 85.00 | ~2GB | ROCm ✅ full GPU |
53
- | gemma4:e4b | 697.91 | 57.34 | ~4GB | ROCm ✅ full GPU |
54
- | gemma4:26b | 157.94 | 10.70 | 12GB + RAM offload | ROCm ⚠️ partial |
55
-
56
- > **Note:** Prefill varies between runs due to KV cache warming.
57
- > Decode is the number that matters for user experience — very consistent.
58
-
59
- ### Platform comparison (decode ~4B models)
60
-
61
- | Hardware | Decode tok/s | Price (approx USD) |
62
- |----------|--------------|--------------------|
63
- | **RX 6700 XT + ROCm (this repo)** | **57-85** | existing hardware |
64
- | Mac Mini M4 16GB | 25-40 | ~$600 new |
65
- | RTX 3070 12GB | 50-80 | ~$500 used |
66
-
67
- **The RX 6700 XT doubles the Mac Mini M4 16GB in decode speed for models that fit in VRAM.**
68
-
69
- ### The gfx1031 problem
70
-
71
- The RX 6700 XT reports `gfx1031` architecture, but ROCm doesn't officially support it.
72
- Without the override, Ollama falls back to CPU (~3-5 tok/s).
73
-
74
- **Fix:**
75
- ```bash
76
- sudo mkdir -p /etc/systemd/system/ollama.service.d
77
-
78
- sudo tee /etc/systemd/system/ollama.service.d/rocm-fix.conf << 'CONF'
79
- [Service]
80
- Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"
81
- Environment="ROCR_VISIBLE_DEVICES=0"
82
- CONF
83
-
84
- sudo systemctl daemon-reload
85
- sudo systemctl restart ollama
86
- ```
87
-
88
- **Verify it worked:**
89
- ```bash
90
- journalctl -u ollama -n 10 --no-pager | grep "AMD Radeon"
91
- # Expected output:
92
- # description="AMD Radeon RX 6700 XT" total="12.0 GiB" available="11.1 GiB"
93
- ```
94
-
95
- ### Full setup
96
-
97
- ```bash
98
- # 1. Install Ollama
99
- curl -fsSL https://ollama.com/install.sh | sh
100
-
101
- # 2. Apply ROCm fix (see above)
102
-
103
- # 3. Pull models
104
- ollama pull gemma4:e2b # 2B - fast, fits easily
105
- ollama pull gemma4:e4b # 4B - great quality, fits in VRAM
106
- ollama pull gemma4:26b # 26B - requires RAM offload on 12GB
107
-
108
- # 4. Run
109
- ollama run gemma4:e4b "Hello, what is MLOps?"
110
- ```
111
-
112
- ### Run the benchmark yourself
113
-
114
- ```bash
115
- chmod +x bench_gemma4.sh
116
- ./bench_gemma4.sh
117
- ```
118
-
119
- Runs 3 iterations per model, saves results to `gemma4_benchmark_results.md`.
120
-
121
- ---
122
-
123
- ## Versión en Español
124
-
125
- ### Por qué esto importa
126
-
127
- La mayoría de guías de AI local asumen:
128
- - GPU NVIDIA (CUDA)
129
- - Apple Silicon (macOS)
130
- - Presupuesto de $1,500 USD o más
131
-
132
- 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.
133
-
134
- ### Hardware
135
-
136
- | Componente | Detalle |
137
- |------------|---------|
138
- | GPU | AMD RX 6700 XT 12GB (gfx1031) |
139
- | CPU | AMD Ryzen 5 5600G |
140
- | RAM | 16GB |
141
- | OS | Pop!_OS 24.04 LTS |
142
- | Ollama | 0.20.2 |
143
- | Override ROCm | HSA_OVERRIDE_GFX_VERSION=10.3.0 |
144
-
145
- ### Resultados (promedio de 3 corridas)
146
-
147
- | Modelo | Prefill (tok/s) | Decode (tok/s) | VRAM | Estado |
148
- |--------|----------------|----------------|------|--------|
149
- | gemma4:e2b | 1022.39 | 85.00 | ~2GB | ROCm ✅ GPU completa |
150
- | gemma4:e4b | 697.91 | 57.34 | ~4GB | ROCm ✅ GPU completa |
151
- | gemma4:26b | 157.94 | 10.70 | 12GB + offload a RAM | ROCm ⚠️ parcial |
152
-
153
- > **Nota:** El prefill varía entre corridas por el KV cache warming.
154
- > El decode es el número relevante para experiencia de usuario y es muy consistente.
155
-
156
- ### Comparativa vs otras plataformas (decode modelos ~4B)
157
-
158
- | Hardware | Decode tok/s | Precio aprox MXN |
159
- |----------|-------------|-----------------|
160
- | **RX 6700 XT + ROCm (este repo)** | **57-85** | hardware existente |
161
- | Mac Mini M4 16GB | 25-40 | ~$23,000 MXN nuevo |
162
- | RTX 3070 12GB | 50-80 | ~$10,000 MXN usado |
163
-
164
- **La RX 6700 XT dobla en velocidad al Mac Mini M4 16GB en modelos que caben en VRAM.**
165
-
166
- ### El problema del gfx1031
167
-
168
- La RX 6700 XT reporta arquitectura `gfx1031` pero ROCm no la tiene en su lista de soporte oficial.
169
- Sin el override, Ollama cae a CPU (~3-5 tok/s).
170
-
171
- **Fix:**
172
- ```bash
173
- sudo mkdir -p /etc/systemd/system/ollama.service.d
174
-
175
- sudo tee /etc/systemd/system/ollama.service.d/rocm-fix.conf << 'CONF'
176
- [Service]
177
- Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"
178
- Environment="ROCR_VISIBLE_DEVICES=0"
179
- CONF
180
-
181
- sudo systemctl daemon-reload
182
- sudo systemctl restart ollama
183
- ```
184
-
185
- **Verificar que funcionó:**
186
- ```bash
187
- journalctl -u ollama -n 10 --no-pager | grep "AMD Radeon"
188
- # Output esperado:
189
- # description="AMD Radeon RX 6700 XT" total="12.0 GiB" available="11.1 GiB"
190
- ```
191
-
192
- ### Setup completo
193
-
194
- ```bash
195
- # 1. Instalar Ollama
196
- curl -fsSL https://ollama.com/install.sh | sh
197
-
198
- # 2. Aplicar fix de ROCm (ver arriba)
199
-
200
- # 3. Descargar modelos
201
- ollama pull gemma4:e2b # 2B - rápido, cabe fácil
202
- ollama pull gemma4:e4b # 4B - buena calidad, cabe en VRAM
203
- ollama pull gemma4:26b # 26B - requiere offload a RAM en 12GB
204
-
205
- # 4. Correr
206
- ollama run gemma4:e4b "Hola, explica qué es MLOps"
207
- ```
208
-
209
- ### Correr el benchmark tú mismo
210
-
211
- ```bash
212
- chmod +x bench_gemma4.sh
213
- ./bench_gemma4.sh
214
- ```
215
-
216
- Corre 3 iteraciones por modelo y guarda los resultados en `gemma4_benchmark_results.md`.
217
-
218
- ### Modelos probados
219
-
220
- | Modelo | Parámetros | Tipo | ¿Cabe en 12GB? |
221
- |--------|-----------|------|----------------|
222
- | gemma4:e2b | 2B | Dense | ✅ GPU completa |
223
- | gemma4:e4b | 4B | Dense | ✅ GPU completa |
224
- | gemma4:26b | 26B (4B activos, MoE) | MoE | ⚠️ offload parcial |
225
-
226
- ---
227
-
228
- ## About / Sobre este repo
229
 
230
- Made in CDMX with second-hand hardware and free software.
231
- Hecho en CDMX con hardware de segunda mano y software libre.
232
 
233
- *By [Positronica Labs](https://github.com/G10hdz) CDMX, México*
234
 
235
  [![GitHub](https://img.shields.io/badge/GitHub-G10hdz-black?logo=github)](https://github.com/G10hdz)
236
  [![LinkedIn](https://img.shields.io/badge/LinkedIn-Mayte%20Giovanna-blue?logo=linkedin)](https://www.linkedin.com/in/mayte-giovanna-hernandez-rios/)
 
3
  - benchmark
4
  - amd
5
  - rocm
 
6
  - local-llm
7
  - linux
 
 
8
  language:
9
  - es
10
  - en
11
  license: apache-2.0
12
  ---
13
 
14
+ # AMD ROCm Benchmarks Positronica Labs
15
 
16
+ Benchmarks de modelos locales corriendo en GPU AMD con ROCm en Linux.
17
+ Para la comunidad LatAm que corre AI sin Apple ni NVIDIA.
18
 
19
+ ## Benchmarks disponibles
 
20
 
21
+ | Carpeta | Modelos | Descripción |
22
+ |---------|---------|-------------|
23
+ | [gemma4/](gemma4/README.md) | Gemma 4 E2B, E4B, 26B | Gemma 4 día 3 del lanzamiento |
24
+ | [all_models/](all_models/README.md) | 11 modelos | Master benchmark con guía de uso |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25
 
26
+ ## Hardware base
 
27
 
28
+ AMD RX 6700 XT 12GB + Ryzen 5 5600G + 16GB RAM + Pop!_OS 24.04
29
 
30
  [![GitHub](https://img.shields.io/badge/GitHub-G10hdz-black?logo=github)](https://github.com/G10hdz)
31
  [![LinkedIn](https://img.shields.io/badge/LinkedIn-Mayte%20Giovanna-blue?logo=linkedin)](https://www.linkedin.com/in/mayte-giovanna-hernandez-rios/)