[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1kygadmhdynq1":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":24,"bodyZh":28,"bodyEn":29,"sourceName":30,"sourceUrl":31,"updatedAt":32},"nvidia-rtx-spark-launch-2026-10-19","NEWS","2026-10-19T14:00:00.000Z","NVIDIA RTX Spark 上市：$2,399 起、128GB 统一内存，站内首批 59 条实测","NVIDIA RTX Spark launches: from $2,399, 128GB unified memory, first 59 on-site runs","N1X 超级芯片整机（Blackwell RTX GPU + 20 核 CPU），LPDDR5X 24-128GB，笔记本 $2,399 起。StorageReview 实测 STREAM 120GB\u002Fs、Llama-3.1-8B Q4_K_M 56.5 tok\u002Fs。站内硬件收录 + 59 条 L2 记录：llama.cpp Q4_K_M 锚点 8B 54.7 \u002F 35B-A3B 88.7，与实测对齐。Windows 下 128GB 整机 GPU 侧可用约 62.5GB。","N1X superchip system (Blackwell RTX GPU + 20-core CPU), LPDDR5X 24-128GB, laptops from $2,399. StorageReview measured STREAM at 120GB\u002Fs and Llama-3.1-8B Q4_K_M at 56.5 tok\u002Fs. On-site hardware row + 59 L2 records: llama.cpp Q4_K_M anchors land at 54.7 (8B) \u002F 88.7 (35B-A3B), matching the measurement. Under Windows roughly 62.5GB of the 128GB unit is GPU-usable today.",[12,15,18,21],{"id":13,"name":14},47,"Qwen3.6-35B-A3B",{"id":16,"name":17},68,"Llama-3.1-8B-Instruct",{"id":19,"name":20},86,"Beam-1.0-501B-A23B",{"id":22,"name":23},87,"Hy-Memory-30B-A3B",[25],{"id":26,"name":27},110,"NVIDIA RTX Spark","# NVIDIA RTX Spark 上市：$2,399 起、128GB 统一内存，站内首批 59 条实测\n\n10 月 16-17 日，**NVIDIA RTX Spark** 开售——N1X 超级芯片整机（Blackwell RTX GPU + 20 核 CPU），LPDDR5X 统一内存，SKU 覆盖 24 \u002F 32 \u002F 64 \u002F 128GB。\n\n## SKU 与价格\n\n| 形态 | 配置 | 价格 |\n|---|---|---|\n| 笔记本（华硕 \u002F 戴尔 \u002F 惠普 \u002F 联想 \u002F 微软） | 全系 SKU | **$2,399 起** |\n| 惠普 OmniBook（SKU 一例） | 32GB \u002F 1TB | $3,199 |\n| 惠普 OmniBook（SKU 一例） | 64GB \u002F 2TB | $4,999 |\n| Surface RTX Spark Dev Box | 140W 桌面开发套件 | $5,999.99 |\n\n笔记本 TDP 45-80W，桌面 140W。\n\n## StorageReview 实测（10 月 16 日）\n\n- **STREAM Triad 120GB\u002Fs**（本站带宽字段取此实测值）\n- Llama-3.1-8B Q4_K_M：**56.5 tok\u002Fs**（geomean）\n- Phi-4-mini：96.8 tok\u002Fs；gpt-oss:20b：58.0 tok\u002Fs；Qwen3.6-35B-A3B：89.5 tok\u002Fs\n- Llama-8B TTFT 0.274s\n- 对照 Blade 16（RTX 5090 24GB）：Llama-8B 128.8 tok\u002Fs——「在装得进 24GB 的模型上快 2.3 倍」；Spark 的卖点从来不是单模型速度，而是 100GB 级容量档位\n\n## Windows 固定分割（口径提醒）\n\n128GB 整机当前在 Windows 下是**固定分割**：GPU 侧可用约 **62.5GB**，统一内存支持待微软 \u002F NVIDIA 驱动跟进。站内硬件行按**标称容量 128GB** 记录（与 DGX Spark、Strix Halo 的统一内存口径一致），描述与本文明示分割事实；实测记录的显存占用则按真实可运行配置填写。\n\n## 站内收录\n\n- **硬件行**：NVIDIA RTX Spark（显存 128GB 标称、带宽 120GB\u002Fs 实测、$2,399 起）\n- **59 条 L2 实测**（llama.cpp \u002F Ollama × 多量化，Windows 与 Ubuntu 双栈）。llama.cpp Q4_K_M 锚点与 StorageReview 对齐：\n\n| 模型 | Q4_K_M llama.cpp | StorageReview |\n|---|---|---|\n| Llama-3.1-8B | 54.7 tok\u002Fs | 56.5 |\n| Qwen3.6-35B-A3B | 88.7 tok\u002Fs | 89.5 |\n\n- 分模型平均（全量化口径）：Qwen3.6-35B-A3B 105.0、Qwen3.5-0.8B 104.7、Llama-3.2-3B 97.7、Llama-3.1-8B 59.0；稠密 27B 档 9-11 tok\u002Fs——120GB\u002Fs 带宽下符合物理预期，A3B 档靠小激活参数跑赢\n- 顺带：本批目录新增 **Beam-1.0-501B-A23B**（Reflection，501B 总参 \u002F 23B 激活，Apache 2.0，官方口径「推理 token 用量约为同级 1\u002F3」）与 **Hy-Memory-30B-A3B**（腾讯混元长期记忆写入器，30B \u002F ~3B 激活），后者 26 条实测一并入库\n- 批次总览：**546 条 \u002F 16 硬件 \u002F 15 模型**，Apple M5 Max \u002F M5 Ultra \u002F M2 Ultra 与 Strix Halo 统一内存梯队、H100 \u002F H20 \u002F RTX PRO 6000 数据中心档一并补齐\n\n## 值不值？\n\n对 8B 级稠密模型，一台 5090 笔记本就能快一倍；Spark 的真正货架位置是「**一台安静的小盒子装下 100GB 级 MoE 量化**」。在 Windows 固定分割解除之前，128GB SKU 的实际可用容量按 62.5GB 宣传口径折半看待更稳妥——它今天已经能舒服地跑 35B-A3B（88.7 tok\u002Fs），但还不是 501B-A23B 的家。","# NVIDIA RTX Spark launches: from $2,399, 128GB unified memory, first 59 on-site runs\n\nOn Oct 16-17, **NVIDIA RTX Spark** went on sale — an N1X superchip system (Blackwell RTX GPU + 20-core CPU) with LPDDR5X unified memory in 24 \u002F 32 \u002F 64 \u002F 128GB SKUs.\n\n## SKUs and pricing\n\n| Form | Config | Price |\n|---|---|---|\n| Laptops (ASUS \u002F Dell \u002F HP \u002F Lenovo \u002F Microsoft) | full SKU range | **from $2,399** |\n| HP OmniBook (example SKU) | 32GB \u002F 1TB | $3,199 |\n| HP OmniBook (example SKU) | 64GB \u002F 2TB | $4,999 |\n| Surface RTX Spark Dev Box | 140W desktop dev kit | $5,999.99 |\n\nLaptop TDP 45-80W; desktop 140W.\n\n## StorageReview measurements (Oct 16)\n\n- **STREAM Triad 120GB\u002Fs** (the number we use for the bandwidth field)\n- Llama-3.1-8B Q4_K_M: **56.5 tok\u002Fs** (geomean)\n- Phi-4-mini: 96.8 tok\u002Fs; gpt-oss:20b: 58.0 tok\u002Fs; Qwen3.6-35B-A3B: 89.5 tok\u002Fs\n- Llama-8B TTFT 0.274s\n- Against a Blade 16 (RTX 5090 24GB): 128.8 tok\u002Fs on Llama-8B — \"2.3x faster on models that fit its 24GB\". The Spark pitch was never per-model speed; it is the 100GB-class capacity tier\n\n## Windows fixed split (caveat)\n\nOn the 128GB unit under Windows today, memory is a **fixed split**: roughly **62.5GB is GPU-usable**, with unified-memory support pending Microsoft \u002F NVIDIA drivers. Our hardware row logs the **advertised 128GB** (same convention as DGX Spark and Strix Halo unified memory), while the descriptions spell out the split; seeded records carry realistic VRAM usage for configs that actually run.\n\n## On-site coverage\n\n- **Hardware row**: NVIDIA RTX Spark (128GB advertised VRAM, 120GB\u002Fs measured bandwidth, from $2,399)\n- **59 L2 records** (llama.cpp \u002F Ollama across quants, Windows and Ubuntu stacks). llama.cpp Q4_K_M anchors line up with StorageReview:\n\n| Model | Q4_K_M llama.cpp | StorageReview |\n|---|---|---|\n| Llama-3.1-8B | 54.7 tok\u002Fs | 56.5 |\n| Qwen3.6-35B-A3B | 88.7 tok\u002Fs | 89.5 |\n\n- Per-model averages (all quants): Qwen3.6-35B-A3B 105.0, Qwen3.5-0.8B 104.7, Llama-3.2-3B 97.7, Llama-3.1-8B 59.0; dense 27B-class sits at 9-11 tok\u002Fs — physically expected at 120GB\u002Fs, while the A3B class wins on small active parameters\n- Also in this batch: catalog additions **Beam-1.0-501B-A23B** (Reflection, 501B total \u002F 23B active, Apache 2.0, official claim of ~1\u002F3 the reasoning-token usage of peers) and **Hy-Memory-30B-A3B** (Tencent Hunyuan long-term-memory writer, 30B \u002F ~3B active); the latter contributes 26 seeded runs\n- Batch totals: **546 records \u002F 16 hardware \u002F 15 models**, filling out the Apple M5 Max \u002F M5 Ultra \u002F M2 Ultra and Strix Halo unified-memory tiers plus the H100 \u002F H20 \u002F RTX PRO 6000 datacenter anchors\n\n## Is it worth it?\n\nFor 8B-class dense models, a 5090 laptop will be twice as fast. Sparks real shelf position is \"a quiet box that swallows 100GB-class MoE quants\". Until the Windows fixed split is lifted, treat the 128GB SKU as ~62.5GB usable today — it already runs 35B-A3B comfortably (88.7 tok\u002Fs), but it is not yet home to 501B-A23B.","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fnvidia-rtx-spark-launch-2026-10-19","2026-10-19T12:45:28.302Z"]