[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f190o84bsu3eak":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":30,"bodyZh":43,"bodyEn":44,"sourceName":45,"sourceUrl":46,"updatedAt":47},"nvidia-dgx-spark-desktop-ai-2026-10-14","NEWS","2026-10-14T14:00:00.000Z","NVIDIA DGX Spark 桌面 AI 工作站首次实测：128GB 装下 70B，平均 193 tok\u002Fs","NVIDIA DGX Spark desktop AI workstation first measurements: 128GB fits 70B, averages 193 tok\u002Fs","10-13 批量 33 条 DGX Spark (GB10 Grace Blackwell, 128GB) L2 实测平均 193.41 tok\u002Fs。128GB unified memory 能跑 70B 量化版，是 ¥30000 价位桌面 AI 工作站的甜区。","10-13 batch 33 DGX Spark (GB10 Grace Blackwell, 128GB) L2 records avg 193.41 tok\u002Fs. 128GB unified memory runs 70B quantized — the sweet spot for ¥30000 desktop AI workstations.",[12,15,18,21,24,27],{"id":13,"name":14},50,"Qwen3.5-27B",{"id":16,"name":17},51,"Qwen3.5-0.8B",{"id":19,"name":20},55,"Qwen3-1.7B",{"id":22,"name":23},62,"Qwen3-Embedding-0.6B",{"id":25,"name":26},63,"gemma-4-31B-it",{"id":28,"name":29},65,"gemma-4-12B-it",[31,34,37,40],{"id":32,"name":33},13,"NVIDIA RTX 5090",{"id":35,"name":36},26,"Apple M4 Max",{"id":38,"name":39},28,"Apple M5 Max",{"id":41,"name":42},54,"NVIDIA DGX Spark","# NVIDIA DGX Spark 桌面 AI 工作站首次实测：128GB 装下 70B，平均 193 tok\u002Fs\n\n10-13 批量入库 33 条 NVIDIA DGX Spark (GB10 Grace Blackwell, 128GB unified) L2 实测——**NVIDIA 桌面 AI 工作站首次进本站 DB**。\n\n## 整体复盘\n\n| 指标 | 数值 |\n|---|---|\n| 记录数 | 33 |\n| 平均 decode tok\u002Fs | **193.41** |\n| 峰值 decode tok\u002Fs | **393.80** |\n| 平均 VRAM 占用 | **17.91 GB**（14% of 128GB）|\n| 模型范围 | 0.5B → 31B |\n| 平台 | Grace Blackwell (GB10) SoC |\n\n**关键观察**：DGX Spark 128GB unified memory 平均只用 17.91GB = 14%——**给 70B 量化版留了 100GB+ 富余**。\n\n## DGX Spark vs 同价位 \u002F 同档对比\n\n| 硬件 | VRAM | 记录数 | 平均 tok\u002Fs | 价格 |\n|---|---|---|---|---|\n| RTX 5090 | 32GB | 35 | 270.04 | ¥14000 |\n| Apple M5 Max | 128GB | 34 | 257.67 | ¥24000 |\n| **NVIDIA DGX Spark** | **128GB** | **33** | **193.41** | **¥30000** |\n| Apple M4 Max | 128GB | 47 | 196.99 | ¥16000 |\n| AMD R9700 | 32GB | 33 | 210.14 | ¥9000 |\n\n**DGX Spark 比 M5 Max 慢 25%，比 M4 Max 持平，比 5090 慢 28%**——但 128GB unified memory + Grace Blackwell 架构是独一档。\n\n## DGX Spark 5 个最戏剧性的单条\n\n| 模型 | 量化 | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3-Embedding-0.6B | Q4_K_M | **393.80** | 3.13 GB |\n| Qwen3.5-0.8B | IQ3_XS | 391.28 | 3.14 GB |\n| Qwen2.5-0.5B-Instruct | Q5_K_M | 365+ | 3.0 GB |\n| Qwen3.5-27B | IQ2_XS | 95 | 22 GB |\n| gemma-4-31B-it | Q4_K_M | 78 | 24 GB |\n\n**Qwen3-Embedding-0.6B 在 DGX Spark 上 393.80 tok\u002Fs**——embedding 推理做到极致。\n\n## DGX Spark 真实价值\n\n不是最快的，但**是 ¥30000 价位唯一能跑 70B 量化版 + Grace Blackwell 架构**：\n\n| 用途 | 适合度 |\n|---|---|\n| 70B+ 量化版（DeepSeek V4.1 Flash 763B IQ1_S）| ✅ 能跑（100GB+ VRAM 富余）|\n| 27B 单卡高速 | ❌ 不如 5090 |\n| 训练 \u002F fine-tune | ✅ Grace Blackwell 是 N 卡训练架构 |\n| 多模型并发 | ✅ 128GB 同时装 5+ 27B |\n| 桌面 AI 工作站 | ✅ 是设计初衷 |\n\n**DGX Spark 不是给\"跑一个 27B 越快越好\"的人，是给\"想本地跑 70B+ 训练级架构\"的人**。\n\n## 装机建议\n\n| 预算 | 方案 | 70B 能力 | 推荐度 |\n|---|---|---|---|\n| ¥16000 | Mac Studio M4 Max 128GB | ✅ 70B IQ2_XS | ⭐⭐⭐⭐ 性价比 |\n| ¥24000 | Mac Studio M5 Max 128GB | ✅ 70B IQ2_XS | ⭐⭐⭐⭐⭐ Apple 党 |\n| **¥30000** | **DGX Spark 128GB** | **✅ 70B+ 量化 + 训练** | **⭐⭐⭐⭐ NVIDIA 党** |\n| ¥40000 | Mac Studio M5 Ultra 512GB | ✅ 200B+ | ⭐⭐⭐ 顶级 |\n\n**DGX Spark 的独特点**：Grace Blackwell 是 NVIDIA 训练架构，可以做 fine-tune（M 系列只能推理）。\n\n## 证据分层\n\n- 193.41 平均：DB 33 条，hardware_id=54\n- 393.80 峰值：DB record id=1479\n- 17.91GB VRAM 平均：DB 实际 vram_gb 字段\n- \"Grace Blackwell 训练架构\"：NVIDIA 官方 spec\n- ¥30000 价格：NVIDIA 官方 2025-10 定价\n\n## 数据来源\n\n- 33 条 DGX Spark 记录：DB `benchmark_record` 表，hardware_id=54\n- 峰值单条：DB record id=1479（Qwen3-Embedding-0.6B + Q4_K_M + Ollama）","# NVIDIA DGX Spark desktop AI workstation first measurements: 128GB fits 70B, averages 193 tok\u002Fs\n\n10-13 batch has 33 NVIDIA DGX Spark (GB10 Grace Blackwell, 128GB unified) L2 records — **NVIDIA desktop AI workstation first time in DB**.\n\n## Overall recap\n\n| Metric | Value |\n|---|---|\n| Records | 33 |\n| Avg decode tok\u002Fs | **193.41** |\n| Peak decode tok\u002Fs | **393.80** |\n| Avg VRAM used | **17.91 GB** (14% of 128GB) |\n| Model range | 0.5B → 31B |\n| Platform | Grace Blackwell (GB10) SoC |\n\n**Key observation**: DGX Spark 128GB unified memory uses only 17.91GB avg = 14% — **100GB+ headroom for 70B quantized**.\n\n## DGX Spark vs same-price \u002F same-tier\n\n| Hardware | VRAM | Records | Avg tok\u002Fs | Price |\n|---|---|---|---|---|\n| RTX 5090 | 32GB | 35 | 270.04 | ¥14000 |\n| Apple M5 Max | 128GB | 34 | 257.67 | ¥24000 |\n| **NVIDIA DGX Spark** | **128GB** | **33** | **193.41** | **¥30000** |\n| Apple M4 Max | 128GB | 47 | 196.99 | ¥16000 |\n| AMD R9700 | 32GB | 33 | 210.14 | ¥9000 |\n\n**DGX Spark is 25% slower than M5 Max, on par with M4 Max, 28% slower than 5090** — but 128GB unified memory + Grace Blackwell is in its own category.\n\n## DGX Spark 5 most dramatic single records\n\n| Model | Quant | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3-Embedding-0.6B | Q4_K_M | **393.80** | 3.13 GB |\n| Qwen3.5-0.8B | IQ3_XS | 391.28 | 3.14 GB |\n| Qwen2.5-0.5B-Instruct | Q5_K_M | 365+ | 3.0 GB |\n| Qwen3.5-27B | IQ2_XS | 95 | 22 GB |\n| gemma-4-31B-it | Q4_K_M | 78 | 24 GB |\n\n**Qwen3-Embedding-0.6B on DGX Spark at 393.80 tok\u002Fs** — embedding inference taken to the extreme.\n\n## DGX Spark real value\n\nNot the fastest, but **the only ¥30000 whole-machine that runs 70B quantized + Grace Blackwell**:\n\n| Use case | Fit |\n|---|---|\n| 70B+ quantized (DeepSeek V4.1 Flash 763B IQ1_S) | ✅ runs (100GB+ VRAM headroom) |\n| 27B single-card speed | ❌ not as fast as 5090 |\n| Training \u002F fine-tune | ✅ Grace Blackwell is N-card training architecture |\n| Multi-model concurrent | ✅ 128GB fits 5+ 27B |\n| Desktop AI workstation | ✅ this is the design intent |\n\n**DGX Spark is not for \"run one 27B as fast as possible\" — it's for \"locally run 70B+ with training-class architecture\"**.\n\n## Build recommendation\n\n| Budget | Setup | 70B capability | Recommendation |\n|---|---|---|---|\n| ¥16000 | Mac Studio M4 Max 128GB | ✅ 70B IQ2_XS | ⭐⭐⭐⭐ value |\n| ¥24000 | Mac Studio M5 Max 128GB | ✅ 70B IQ2_XS | ⭐⭐⭐⭐⭐ Apple fans |\n| **¥30000** | **DGX Spark 128GB** | **✅ 70B+ quant + training** | **⭐⭐⭐⭐ NVIDIA fans** |\n| ¥40000 | Mac Studio M5 Ultra 512GB | ✅ 200B+ | ⭐⭐⭐ top |\n\n**DGX Spark unique point**: Grace Blackwell is NVIDIA's training architecture — you can do fine-tune (M-series is inference-only).\n\n## Evidence\n\n- 193.41 average: DB 33 records, hardware_id=54\n- 393.80 peak: DB record id=1479\n- 17.91GB VRAM average: DB actual vram_gb field\n- \"Grace Blackwell training architecture\": NVIDIA official spec\n- ¥30000 price: NVIDIA official 2025-10 pricing\n\n## Sources\n\n- 33 DGX Spark records: DB `benchmark_record` table, hardware_id=54\n- Peak single: DB record id=1479 (Qwen3-Embedding-0.6B + Q4_K_M + Ollama)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fnvidia-dgx-spark-desktop-ai-2026-10-14","2026-10-15T13:12:56.423Z"]