趋近智
Radeon PRO W6800 (32GB)
Radeon PRO W6800 (32GB) 提供 32 GB 显存/内存与 512 GB/s 带宽，支持通过 ROCm 与 Vulkan 加速进行本地推理。
总显存
32 GB
可用显存上限
32 GB
显存带宽
512 GB/s
最大推理模型级别
32B 模型 (Q4)
热设计功耗
250W
显存占用随上下文长度变化曲线
随上下文长度增长的显存需求评估。当曲线超过参考线时将导致显存不足（OOM）。
量化精度:
Radeon PRO W6800 (32GB) 最佳运行大模型与速度矩阵
主流开源大模型在 Radeon PRO W6800 (32GB) 上的推理速度、显存兼容性与生成吞吐表现。
排序:
上下文长度:
KV 缓存:
显卡数量:
| 排名 | 模型 | 参数量 | 量化精度 | 预估 TPS | TTFT | 可运行 |
|---|---|---|---|---|---|---|
#36 | Qwen 3.8 27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
#80 | K2 Horizon MoVA 36B A4B | 36B (4B active) | INT4 | 104 tok/s | ~1.6s | |
#86 | Qwen3.5-27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
#93 | Sarvam-30B | 32B (2.4B active) | INT4 | 133 tok/s | ~1.1s | |
#99 | Muse Glimmer 30B | 30B | INT4 Q8 | 24 tok/s 12 tok/s | ~9.3s ~10.9s | |
#102 | Phi-4 Reasoning Plus | 14B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 12 tok/s | ~4.4s ~4.4s ~5.2s | |
#103 | Qwen3.6 35B A3B | 35B (3B active) | INT4 | 120 tok/s | ~1.3s | |
#105 | MiMo V2 Flash | 15B (309B active) | INT4 Q8 | 3 tok/s 1 tok/s | ~91.2s ~91.6s | |
#110 | Gemma 4 26B A4B | 25.2B (3.8B active) | INT4 Q8 | 111 tok/s 74 tok/s | ~1.5s ~1.6s | |
#120 | Qwen3.5-35B-A3B | 35B (3B active) | INT4 | 120 tok/s | ~1.3s | |
#121 | Qwen3.5-9B | 9B | INT4 Q8 FP16 | 68 tok/s 41 tok/s 22 tok/s | ~2.8s ~2.9s ~2.9s | |
#127 | K2 Horizon 7B | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
#130 | Gemma 4 12B | 11.95B | INT4 Q8 FP16 | 39 tok/s 27 tok/s 15 tok/s | ~3.8s ~3.8s ~4.1s | |
#136 | Qwen3.5-4B | 4B | INT4 Q8 FP16 | 122 tok/s 81 tok/s 46 tok/s | ~1.3s ~1.3s ~1.3s | |
#139 | GPT-OSS 20B | 21B (3.6B active) | INT4 Q8 | 54 tok/s 45 tok/s | ~1.4s ~1.5s | |
#140 | GLM-4V | 9B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.8s ~2.9s ~2.9s | |
#145 | NVIDIA Nemotron 3 Nano 30B-A3B | 3.5B (30B active) | INT4 Q8 FP16 | 23 tok/s 13 tok/s 7 tok/s | ~9.0s ~9.0s ~9.1s | |
#150 | Qwen3-30B-A3B | 30B (3B active) | INT4 Q8 | 122 tok/s 80 tok/s | ~1.3s ~1.5s | |
#152 | Granite 4.2 8B | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#157 | Gemma 4 E4B | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#158 | Mistral Large 3 | 41B (675B active) | INT4 | 1 tok/s | ~234.2s | |
#161 | GLM-4 | 32B | INT4 | 17 tok/s | ~9.9s | |
#163 | Gemma 3 27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
#164 | Phi-4 | 14B | INT4 Q8 FP16 | 47 tok/s 28 tok/s 14 tok/s | ~4.4s ~4.4s ~4.8s | |
#166 | Qwen3-32B | 32B | INT4 | 22 tok/s | ~9.9s | |
#168 | DeepSeek-R1 32B | 32B | INT4 Q8 | 23 tok/s 12 tok/s | ~9.9s ~11.6s | |
#169 | Gemma 4 31B | 30.7B | INT4 Q8 | 23 tok/s 12 tok/s | ~9.5s ~11.2s | |
#171 | Magistral Small | 24B | INT4 Q8 | 22 tok/s 14 tok/s | ~7.4s ~8.1s | |
#172 | Qwen2.5-32B | 32B | INT4 | 22 tok/s | ~9.9s | |
#174 | OLMo 3.1 32B Instruct | 32B | INT4 | 17 tok/s | ~9.9s | |
#176 | DeepSeek-R1 14B | 14B | INT4 Q8 FP16 | 48 tok/s 28 tok/s 14 tok/s | ~4.4s ~4.4s ~4.8s | |
#177 | Ministral 3 14B | 14B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 12 tok/s | ~4.4s ~4.4s ~5.2s | |
#180 | Gemma 3 12B | 12B | INT4 Q8 FP16 | 54 tok/s 32 tok/s 17 tok/s | ~3.8s ~3.8s ~3.8s | |
#184 | Mistral-Small-2501 | 24B | INT4 Q8 | 29 tok/s 16 tok/s | ~7.4s ~8.1s | |
#185 | Ministral 3 8B | 8B | INT4 Q8 FP16 | 54 tok/s 38 tok/s 22 tok/s | ~2.5s ~2.5s ~2.6s | |
#186 | Gemma 2 27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
#187 | OLMo 3 7B Think | 7B | INT4 Q8 FP16 | 60 tok/s 42 tok/s 25 tok/s | ~2.2s ~2.2s ~2.3s | |
#188 | Qwen3.5-2B | 2B | INT4 Q8 FP16 | 178 tok/s 130 tok/s 81 tok/s | ~695 ms ~697 ms ~702 ms | |
#189 | MiniCPM5-2B | 2B | INT4 Q8 FP16 | 178 tok/s 130 tok/s 81 tok/s | ~695 ms ~697 ms ~702 ms | |
#190 | Qwen2.5-14B | 14B | INT4 Q8 FP16 | 47 tok/s 28 tok/s 14 tok/s | ~4.4s ~4.4s ~4.8s | |
#191 | OLMo 3.1 32B Think | 32B | INT4 | 17 tok/s | ~9.9s | |
#193 | Gemma 4 E2B | 5.1B | INT4 Q8 FP16 | 104 tok/s 67 tok/s 38 tok/s | ~1.6s ~1.6s ~1.7s | |
#195 | Ministral 3 3B | 3B | INT4 Q8 FP16 | 99 tok/s 77 tok/s 50 tok/s | ~1.0s ~1.0s ~1.0s | |
#196 | DeepSeek-R1 8B | 8B | INT4 Q8 FP16 | 76 tok/s 47 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#198 | Gemma 2 9B | 9B | INT4 Q8 FP16 | 68 tok/s 41 tok/s 22 tok/s | ~2.8s ~2.9s ~2.9s | |
#199 | Nemotron 3.5 Lightning 30B A3B | 30B (3B active) | INT4 Q8 | 122 tok/s 80 tok/s | ~1.3s ~1.5s | |
#200 | Kimi Linear 48B A3B Instruct | 48B (3B active) | INT4 | 37 tok/s | ~1.7s | |
#201 | Qwen3-4B | 4B | INT4 Q8 FP16 | 122 tok/s 81 tok/s 46 tok/s | ~1.3s ~1.3s ~1.3s | |
#202 | Qwen3-14B | 14B | INT4 Q8 FP16 | 47 tok/s 28 tok/s 14 tok/s | ~4.4s ~4.4s ~4.8s | |
#203 | Gemma 3 4B | 4B | INT4 Q8 FP16 | 122 tok/s 81 tok/s 46 tok/s | ~1.3s ~1.3s ~1.3s | |
#204 | DeepSeek-R1 7B | 7B | INT4 Q8 FP16 | 85 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
#205 | OLMo 3 32B Base | 32B | INT4 | 17 tok/s | ~9.9s | |
#206 | Qwen3.5-0.8B | 0.8B | INT4 Q8 FP16 | 258 tok/s 213 tok/s 153 tok/s | ~285 ms ~286 ms ~288 ms | |
#207 | Llama 3.1 8B | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#207 | Qwen3-8B | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#210 | Nemotron 3.5 Lightning | 30B (3B active) | INT4 | 49 tok/s | ~1.3s | |
#211 | Command R | 35B | INT4 | 16 tok/s | ~10.8s | |
#212 | Llama 3 8B | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#213 | Qwen3-1.7B | 1.7B | INT4 Q8 FP16 | 193 tok/s 144 tok/s 92 tok/s | ~595 ms ~597 ms ~601 ms | |
#214 | Gemma 3 270M | 0.27B | INT4 Q8 FP16 | 296 tok/s 275 tok/s 235 tok/s | ~98 ms ~98 ms ~98 ms | |
#216 | Qwen3-0.6B | 0.6B | INT4 Q8 FP16 | 280 tok/s 239 tok/s 180 tok/s | ~212 ms ~213 ms ~214 ms | |
#217 | Phi-4-Mini | 3.8B | INT4 Q8 FP16 | 126 tok/s 84 tok/s 48 tok/s | ~1.2s ~1.2s ~1.3s | |
#218 | Qwen2.5-7B | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
#219 | Mixtral-8x7B-v0.1 | 46.7B (7B active) | INT4 | 70 tok/s | ~2.8s | |
#220 | Ministral-8B-2410 | 8B | INT4 Q8 FP16 | 75 tok/s 46 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
#221 | DeepSeek-R1 1.5B | 1.5B | INT4 Q8 FP16 | 204 tok/s 155 tok/s 101 tok/s | ~527 ms ~528 ms ~532 ms | |
#222 | Llama 3.2 3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
#223 | Phi-3-mini | 3.8B | INT4 Q8 FP16 | 126 tok/s 84 tok/s 48 tok/s | ~1.2s ~1.2s ~1.3s | |
#224 | Gemma 3n E2B IT | 6B (2B active) | INT4 Q8 FP16 | 99 tok/s 84 tok/s 61 tok/s | ~746 ms ~748 ms ~752 ms | |
#225 | Gemma 2 2B | 2B | INT4 Q8 FP16 | 178 tok/s 130 tok/s 81 tok/s | ~695 ms ~697 ms ~702 ms | |
#226 | Mistral-7B-Instruct-v0.1 | 7.3B | INT4 Q8 FP16 | 80 tok/s 50 tok/s 27 tok/s | ~2.3s ~2.3s ~2.3s | |
#227 | Llama 3.2 1B | 1B | INT4 Q8 FP16 | 238 tok/s 191 tok/s 133 tok/s | ~361 ms ~362 ms ~364 ms | |
#228 | Phi-3-medium | 14B | INT4 Q8 FP16 | 47 tok/s 28 tok/s 14 tok/s | ~4.4s ~4.4s ~4.8s | |
#229 | Yi-34B | 34B | INT4 | 17 tok/s | ~10.5s | |
#230 | OLMo 3 7B Instruct | 7B | INT4 Q8 FP16 | 60 tok/s 42 tok/s 25 tok/s | ~2.2s ~2.2s ~2.3s | |
#231 | Qwen2-7B | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
#232 | Phi-3-small | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
#233 | Mistral-7B-Instruct-v0.2 | 7.3B | INT4 Q8 FP16 | 80 tok/s 50 tok/s 27 tok/s | ~2.3s ~2.3s ~2.3s | |
#234 | Gemma 3 1B | 1B | INT4 Q8 FP16 | 238 tok/s 191 tok/s 133 tok/s | ~361 ms ~362 ms ~364 ms | |
#236 | Gemma 1 7B | 7B | INT4 Q8 FP16 | 60 tok/s 42 tok/s 25 tok/s | ~2.2s ~2.2s ~2.3s | |
#237 | Gemma 1 2B | 2B | INT4 Q8 FP16 | 189 tok/s 135 tok/s 83 tok/s | ~695 ms ~697 ms ~701 ms | |
#238 | ChatGLM3-6B | 6B | INT4 Q8 FP16 | 66 tok/s 48 tok/s 29 tok/s | ~1.9s ~1.9s ~1.9s | |
#239 | ChatGLM2-6B | 6B | INT4 Q8 FP16 | 66 tok/s 48 tok/s 29 tok/s | ~1.9s ~1.9s ~1.9s | |
#240 | ChatGLM-6B | 6B | INT4 Q8 FP16 | 66 tok/s 48 tok/s 29 tok/s | ~1.9s ~1.9s ~1.9s | |
未上榜模型（按推理速度排序） | ||||||
- | DINOv3 ViT-L/16 | 0.3B | INT4 Q8 FP16 | 320 tok/s 291 tok/s 242 tok/s | ~108 ms ~108 ms ~109 ms | |
- | Qwen2-0.5B | 0.5B | INT4 Q8 FP16 | 293 tok/s 254 tok/s 197 tok/s | ~176 ms ~177 ms ~178 ms | |
- | Qwen2.5-0.5B | 0.5B | INT4 Q8 FP16 | 293 tok/s 254 tok/s 197 tok/s | ~176 ms ~177 ms ~178 ms | |
- | ERNIE-4.5-0.3B | 0.3B | INT4 Q8 FP16 | 288 tok/s 266 tok/s 225 tok/s | ~108 ms ~108 ms ~109 ms | |
- | ERNIE-4.5-0.3B-Base | 0.3B | INT4 Q8 FP16 | 288 tok/s 266 tok/s 225 tok/s | ~108 ms ~108 ms ~109 ms | |
- | Liquid D1-Omni 600M | 0.6B | INT4 Q8 FP16 | 280 tok/s 239 tok/s 180 tok/s | ~212 ms ~213 ms ~214 ms | |
- | Falcon-1B | 1B | INT4 Q8 FP16 | 249 tok/s 198 tok/s 136 tok/s | ~361 ms ~362 ms ~364 ms | |
- | Falcon3-1B | 1B | INT4 Q8 FP16 | 238 tok/s 191 tok/s 133 tok/s | ~361 ms ~362 ms ~364 ms | |
- | MiniCPM-V 4.7 1B | 1B | INT4 Q8 FP16 | 238 tok/s 191 tok/s 133 tok/s | ~361 ms ~362 ms ~364 ms | |
- | Qwen2-1.5B | 1.5B | INT4 Q8 FP16 | 204 tok/s 155 tok/s 101 tok/s | ~527 ms ~528 ms ~532 ms | |
- | Qwen2.5-1.5B | 1.5B | INT4 Q8 FP16 | 204 tok/s 155 tok/s 101 tok/s | ~527 ms ~528 ms ~532 ms | |
- | Spark X2.5 1.7B | 1.7B | INT4 Q8 FP16 | 193 tok/s 144 tok/s 92 tok/s | ~595 ms ~597 ms ~601 ms | |
- | Hy-MT2 1.8B | 1.8B | INT4 Q8 FP16 | 188 tok/s 139 tok/s 88 tok/s | ~626 ms ~628 ms ~632 ms | |
- | Falcon-3B | 3B | INT4 Q8 FP16 | 153 tok/s 103 tok/s 60 tok/s | ~999 ms ~1.0s ~1.0s | |
- | CroissantLLM Base | 1.3B | INT4 Q8 FP16 | 150 tok/s 128 tok/s 92 tok/s | ~461 ms ~462 ms ~465 ms | |
- | Phi-1 | 1.3B | INT4 Q8 FP16 | 150 tok/s 128 tok/s 92 tok/s | ~461 ms ~462 ms ~465 ms | |
- | Phi-1.5 | 1.3B | INT4 Q8 FP16 | 150 tok/s 128 tok/s 92 tok/s | ~461 ms ~462 ms ~465 ms | |
- | DeepSeek-R1 3B | 3B | INT4 Q8 FP16 | 147 tok/s 101 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | Mellum 2.1 12B A2.5B Thinking | 12B (2.5B active) | INT4 Q8 FP16 | 147 tok/s 107 tok/s 66 tok/s | ~976 ms ~978 ms ~984 ms | |
- | Falcon3-3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | LFM2-VL 3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | Liquid D1 3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | Liquid D1-VL 3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | Liquid LFM2-VL 3B D1 | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | Qwen2.5-3B | 3B | INT4 Q8 FP16 | 145 tok/s 100 tok/s 59 tok/s | ~999 ms ~1.0s ~1.0s | |
- | LFM2.5 24B A2B | 24B (2.5B active) | INT4 Q8 | 136 tok/s 96 tok/s | ~1.1s ~1.2s | |
- | ERNIE-4.5-21B-A3B | 21B (3B active) | INT4 Q8 | 127 tok/s 92 tok/s | ~1.2s ~1.2s | |
- | ERNIE-4.5-21B-A3B-Base | 21B (3B active) | INT4 Q8 | 127 tok/s 92 tok/s | ~1.2s ~1.2s | |
- | ERNIE-4.5-VL-28B-A3B | 28B (3B active) | INT4 Q8 | 123 tok/s 85 tok/s | ~1.3s ~1.4s | |
- | ERNIE-4.5-VL-28B-A3B-Base | 28B (3B active) | INT4 Q8 | 123 tok/s 85 tok/s | ~1.3s ~1.4s | |
- | FrogNano 4B | 4B | INT4 Q8 FP16 | 122 tok/s 81 tok/s 46 tok/s | ~1.3s ~1.3s ~1.3s | |
- | Hy-Memory 30B A3B | 30B (3B active) | INT4 Q8 | 122 tok/s 80 tok/s | ~1.3s ~1.5s | |
- | Hy-MT2 30B A3B | 30B (3B active) | INT4 Q8 | 122 tok/s 80 tok/s | ~1.3s ~1.5s | |
- | Spark X2.5 4B | 4B | INT4 Q8 FP16 | 122 tok/s 81 tok/s 46 tok/s | ~1.3s ~1.3s ~1.3s | |
- | MiniCPM-V 4.7 35B A3B | 35B (3B active) | INT4 | 120 tok/s | ~1.3s | |
- | EmbeddingGemma 2 | 2B | INT4 Q8 FP16 | 117 tok/s 96 tok/s 67 tok/s | ~698 ms ~700 ms ~704 ms | |
- | Phi-2 | 2.7B | INT4 Q8 FP16 | 104 tok/s 82 tok/s 54 tok/s | ~911 ms ~914 ms ~920 ms | |
- | MaLLaM-3B | 3B | INT4 Q8 FP16 | 99 tok/s 77 tok/s 50 tok/s | ~1.0s ~1.0s ~1.0s | |
- | SmolLM3 3B | 3B | INT4 Q8 FP16 | 99 tok/s 77 tok/s 50 tok/s | ~1.0s ~1.0s ~1.0s | |
- | Falcon-7B | 7B | INT4 Q8 FP16 | 86 tok/s 53 tok/s 29 tok/s | ~2.2s ~2.2s ~2.2s | |
- | Falcon3-7B | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
- | Hy-MT2 7B | 7B | INT4 Q8 FP16 | 83 tok/s 52 tok/s 28 tok/s | ~2.2s ~2.2s ~2.2s | |
- | Mistral-7B-v0.1 | 7.3B | INT4 Q8 FP16 | 80 tok/s 50 tok/s 27 tok/s | ~2.3s ~2.3s ~2.3s | |
- | Youtu Parsing Omni | 8B | INT4 Q8 FP16 | 76 tok/s 47 tok/s 25 tok/s | ~2.5s ~2.5s ~2.6s | |
- | LensVLM 9B | 9B | INT4 Q8 FP16 | 68 tok/s 41 tok/s 22 tok/s | ~2.8s ~2.9s ~2.9s | |
- | MiMo V2.6 Distill Qwen 9B | 9B | INT4 Q8 FP16 | 68 tok/s 41 tok/s 22 tok/s | ~2.8s ~2.9s ~2.9s | |
- | ChatGLM3-6B-32K | 6B | INT4 Q8 FP16 | 66 tok/s 48 tok/s 29 tok/s | ~1.9s ~1.9s ~1.9s | |
- | Yi-6B | 6B | INT4 Q8 FP16 | 66 tok/s 48 tok/s 29 tok/s | ~1.9s ~1.9s ~1.9s | |
- | Falcon3-10B | 10B | INT4 Q8 FP16 | 62 tok/s 38 tok/s 20 tok/s | ~3.2s ~3.2s ~3.2s | |
- | Kimi-VL-A3B-Instruct | 16B (3B active) | INT4 Q8 | 62 tok/s 54 tok/s | ~1.2s ~1.2s | |
- | Falcon2-11B | 11B | INT4 Q8 FP16 | 60 tok/s 35 tok/s 19 tok/s | ~3.5s ~3.5s ~3.5s | |
- | Hunyuan Lite | 7B | INT4 Q8 FP16 | 60 tok/s 42 tok/s 25 tok/s | ~2.2s ~2.2s ~2.3s | |
- | OLMo 3 7B Base | 7B | INT4 Q8 FP16 | 60 tok/s 42 tok/s 25 tok/s | ~2.2s ~2.2s ~2.3s | |
- | SEA-LION-7B | 7.1B | INT4 Q8 FP16 | 59 tok/s 42 tok/s 25 tok/s | ~2.3s ~2.3s ~2.3s | |
- | SEA-LION-7B-Instruct | 7.1B | INT4 Q8 FP16 | 59 tok/s 42 tok/s 25 tok/s | ~2.3s ~2.3s ~2.3s | |
- | Sahabat-AI-Llama3-8B-Instruct | 8B | INT4 Q8 FP16 | 54 tok/s 38 tok/s 22 tok/s | ~2.5s ~2.5s ~2.6s | |
- | Typhoon-2-8B | 8B | INT4 Q8 FP16 | 54 tok/s 38 tok/s 22 tok/s | ~2.5s ~2.5s ~2.6s | |
- | GLM-4-9B | 9B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.8s ~2.9s ~2.9s | |
- | GLM-4-9B-Chat | 9B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.8s ~2.9s ~2.9s | |
- | GLM-4-9B-Chat-1M | 9B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.8s ~2.9s ~2.9s | |
- | Sahabat-AI-Gemma2-9B | 9.2B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.9s ~2.9s ~2.9s | |
- | Sahabat-AI-Gemma2-9B-Instruct | 9.2B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.9s ~2.9s ~2.9s | |
- | Yi-9B | 9B | INT4 Q8 FP16 | 49 tok/s 34 tok/s 20 tok/s | ~2.8s ~2.9s ~2.9s | |
- | GreenMind-14B-R1 | 14B | INT4 Q8 FP16 | 34 tok/s 23 tok/s 12 tok/s | ~4.4s ~4.4s ~5.2s | |
- | AREX-2 | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
- | PPLX Decider v1 27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
- | PPLX Decider v1.1 27B | 27B | INT4 Q8 | 26 tok/s 14 tok/s | ~8.3s ~9.1s | |
- | Agnes 3.0 Flash | 33B | INT4 | 22 tok/s | ~10.1s | |
- | Falcon-40B | 40B | INT4 | 19 tok/s | ~12.3s | |
推理能力评估
不同量化精度下可运行的最大模型参数级别。
上下文长度:
4位量化
32B 模型
8位精度
14B 模型
16位全精度
14B 模型
微调能力评估
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
32B 模型
LoRA
14B 模型
全参数微调
3B 模型
负载与选型建议
Radeon PRO W6800 (32GB) 上的最佳量化精度与运行规格建议。
最佳推理区间
适合在全精度或 Q8 精度下运行 14B 模型，以及在 Q4 精度下搭配最高 32k 上下文运行 32B 模型。
显存带宽与推理速率
具备 512 GB/s 聚合带宽，单批次推理处于显存带宽受限模式。8B Q4 模型预估生成速度约 114 tok/s，70B Q4 模型预估生成速度约 13 tok/s。