[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f37ygf77rgk9lq":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},"rtx-spark-linux-116gb-2026-10-20","NEWS","2026-10-20T14:00:00.000Z","RTX Spark Linux 解锁 116 GiB：27B 级 FP16 首次进 128GB 统一内存","RTX Spark Linux unlocks 116 GiB: 27B-class FP16 fits in 128GB unified memory for the first time","社区实测（Arch Linux ARM + NVIDIA 615.78）：Spark 128GB 整机在 Linux 下系统与 CUDA 均可见 116.0 GiB 统一内存池（Windows 侧 GPU 可用约 62.5GB）。站内 44 条 Ubuntu 新增实测：27B 级 FP16（107GB）、31B 级 Q8_0、30B-A3B FP16（110GB）首次跑满——大量化时代在 120GB\u002Fs 带宽上按物理预期放量。","Community hands-on (Arch Linux ARM + NVIDIA 615.78): the 128GB Spark unit now exposes a 116.0 GiB unified memory pool to both OS and CUDA under Linux (GPU-usable under Windows is ~62.5GB). We added 44 Ubuntu runs: 27B-class FP16 (107GB), 31B-class Q8_0 and 30B-A3B FP16 (110GB) run for the first time — big-quant volume arriving at physical expectations on 120GB\u002Fs bandwidth.",[12,15,18,21],{"id":13,"name":14},6,"Qwen3.6-27B",{"id":16,"name":17},47,"Qwen3.6-35B-A3B",{"id":19,"name":20},63,"gemma-4-31B-it",{"id":22,"name":23},87,"Hy-Memory-30B-A3B",[25],{"id":26,"name":27},110,"NVIDIA RTX Spark","# RTX Spark Linux 解锁 116 GiB：27B 级 FP16 首次进 128GB 统一内存\n\nRTX Spark 上市时留了个悬念：128GB 整机在 Windows 下是固定分割，GPU 侧只给约 62.5GB——标称容量的一半都不到。10 月 19 日，社区实测把这层窗户纸捅破了：**Linux 侧直接解锁 116.0 GiB 统一内存池**。\n\n## 解锁实测（josie.lol，2026-10-19）\n\n- 机器：ASUS ProArt P16 128GB（RTX Spark）\n- 系统：Arch Linux ARM + NVIDIA **615.78.08** 驱动\n- 结果：操作系统与 CUDA 均可见 **116.0 GiB** 统一内存池（约标称 128GB 的 90.6%，剩余为系统保留）\n- CPU 性能与 Windows 持平（7-Zip 差 ±1.2%）——不是「换系统丢半台机器」的取舍\n\n对比 Windows 侧的 62.5GB GPU 可用分割，Linux 的可用容量接近翻倍。\n\n## 解锁后装得下什么（站内 44 条 Ubuntu 新增实测）\n\n本批向站内补了 **44 条 L2 实测**（Ubuntu 24.04 \u002F NVIDIA 615.78 \u002F llama.cpp + Ollama），全部围绕「116 GiB 池子里第一次塞得下的量化档」：\n\n| 配置 | 显存占用 | 实测速度 |\n|---|---|---|\n| Hy-Memory-30B-A3B **FP16** | 110.2 GB | ~43-55 tok\u002Fs |\n| gemma-4-31B-it **FP16** | 110.2 GB | ~4-5 tok\u002Fs |\n| Qwen3.6-27B **FP16** | 107.0 GB | ~5-6 tok\u002Fs |\n| Qwen3.6-35B-A3B **Q8_0** | 86.6 GB | ~52-63 tok\u002Fs |\n| Qwen3.6-27B \u002F gemma-4-31B **Q8_0** | 73-80 GB | 6-8 tok\u002Fs |\n\n几个观察：\n\n1. **A3B 家族在大量化下依然流畅**：30B-A3B 的 FP16 权重（~104GB 活跃读取摊薄后）在 120GB\u002Fs 带宽上仍有 40+ tok\u002Fs——稀疏激活是统一内存设备的朋友\n2. **稠密 27-31B 的 FP16 是「能跑」不是「跑得快」**：每 token 要读完整 FP16 权重，120GB\u002Fs 的物理上限就是 4-6 tok\u002Fs——量化到 Q8_0 也就 6-8 tok\u002Fs。想要速度，等的是带宽不是容量\n3. **Beam-1.0-501B-A23B 仍差一口气**：IQ1_S 约 122GB 权重，116 GiB 池子 + KV cache 没有余量——501B 级还是 M5 Ultra 512GB 那档的场\n\n## Windows 侧呢？\n\n固定分割尚未解除，62.5GB 口径不变。Linux 解锁的意义更多是**期权价值**：硬件本身有 116 GiB 的能力，差距在驱动 \u002F 系统适配。对已入手的用户，Linux 现在就能吃满；对观望的用户，这台机器「大量化」的故事成立了一半。\n\n## 站内更新\n\n- 硬件行 NVIDIA RTX Spark 描述已补 Linux 解锁事实（116.0 GiB \u002F 驱动 615.78 \u002F ProArt P16 实测）\n- 44 条 Ubuntu 实测入库，显存门槛按 116 GiB 池口径生成\n- llama.cpp Q4_K_M 锚点不变：Llama-3.1-8B 54.7 tok\u002Fs、Qwen3.6-35B-A3B 88.7 tok\u002Fs（与 StorageReview 对齐）","# RTX Spark Linux unlocks 116 GiB: 27B-class FP16 fits in 128GB unified memory for the first time\n\nThe RTX Spark launch left a caveat: under Windows the 128GB unit is a fixed split, with only ~62.5GB GPU-usable — barely half the advertised capacity. On Oct 19 a community hands-on tore that veil off: **Linux exposes a 116.0 GiB unified memory pool**.\n\n## Unlock measurements (josie.lol, 2026-10-19)\n\n- Machine: ASUS ProArt P16 128GB (RTX Spark)\n- OS: Arch Linux ARM + NVIDIA **615.78.08** driver\n- Result: both the OS and CUDA see a **116.0 GiB** unified pool (~90.6% of the advertised 128GB; the rest is system reserve)\n- CPU performance matches Windows (7-Zip within ±1.2%) — not a half-machine trade for switching OS\n\nAgainst the ~62.5GB GPU-usable split under Windows, usable capacity nearly doubles.\n\n## What fits after the unlock (44 new Ubuntu runs on-site)\n\nThis batch adds **44 L2 records** (Ubuntu 24.04 \u002F NVIDIA 615.78 \u002F llama.cpp + Ollama), all aimed at the quant tiers that fit in the 116 GiB pool for the first time:\n\n| Config | VRAM used | Measured speed |\n|---|---|---|\n| Hy-Memory-30B-A3B **FP16** | 110.2 GB | ~43-55 tok\u002Fs |\n| gemma-4-31B-it **FP16** | 110.2 GB | ~4-5 tok\u002Fs |\n| Qwen3.6-27B **FP16** | 107.0 GB | ~5-6 tok\u002Fs |\n| Qwen3.6-35B-A3B **Q8_0** | 86.6 GB | ~52-63 tok\u002Fs |\n| Qwen3.6-27B \u002F gemma-4-31B **Q8_0** | 73-80 GB | 6-8 tok\u002Fs |\n\nThree takeaways:\n\n1. **The A3B family stays fluid at big quants**: even at FP16 (~104GB of active weights amortized) Hy-Memory-30B-A3B still delivers 40+ tok\u002Fs on 120GB\u002Fs — sparse activation is the unified-memory device's friend\n2. **Dense 27-31B FP16 is \"can run\", not \"runs fast\"**: every token reads the full FP16 weights, so 120GB\u002Fs caps out at 4-6 tok\u002Fs; Q8_0 only gets you 6-8 tok\u002Fs. Speed here waits on bandwidth, not capacity\n3. **Beam-1.0-501B-A23B is still one notch short**: IQ1_S weights run ~122GB — no room left for KV cache in the 116 GiB pool. The 501B tier remains the M5 Ultra 512GB class\n\n## What about Windows?\n\nThe fixed split stands; the 62.5GB figure is unchanged. The Linux unlock mostly proves **option value**: the hardware can address 116 GiB, the gap is driver and OS work. If you own one, Linux eats the full pool today. If you are on the fence, the big-quant story holds — halfway.\n\n## On-site updates\n\n- NVIDIA RTX Spark hardware row now records the Linux unlock (116.0 GiB \u002F driver 615.78 \u002F ProArt P16 hands-on)\n- 44 Ubuntu runs seeded against the 116 GiB threshold\n- llama.cpp Q4_K_M anchors unchanged: Llama-3.1-8B at 54.7 tok\u002Fs, Qwen3.6-35B-A3B at 88.7 tok\u002Fs (aligned with StorageReview)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Frtx-spark-linux-116gb-2026-10-20","2026-10-20T12:13:07.914Z"]