英伟达 DGX Spark 支持 8 台集群 128GB 版本统一内存最高达 1TB RTX PRO 6000 Blackwell 相关任务提速 40%NVIDIA DGX Spark Supports 8-Unit Clusters; 128GB Unified Memory Tops 1TB with RTX


AI free token/AI免费积分

IT之家 10 月 21 日消息,英伟达今日在柏林 GTC 2026 大会上宣布,DGX Spark 此前兼容最多 4 台设备组建集群,当下已开放手动规格最多 8 台设备组建集群。如果是以 128GB 版本组建,其统一内存容量最高可达 1TB。

英伟达 DGX Spark 支持 8 台集群 128GB 版本统一内存最高达 1TB RTX PRO 6000 Blackwell 相关任务提速 40%

同时,英伟达还发布了本地 AI 平台更新,并宣布与 llama.cpp、vLLM 开源社区合作优化 AI 推理表现,在特定工作负载下,RTX PRO 6000 Blackwell 的推理表现最高提高 40%,DGX Spark 和 DGX Station 也分别取得最高 30% 和 20% 的表现提高。

英伟达 DGX Spark 支持 8 台集群 128GB 版本统一内存最高达 1TB RTX PRO 6000 Blackwell 相关任务提速 40%

英伟达表示,4,999 美元(IT之家注:现汇率约合 33,450 元人民币)的 DGX Spark 的 64GB 版本将于 10 月 23 日开始借助合作伙伴陆续发售。针对集群部署,英伟达 Sync Model Launcher 计划于本月晚些时候上线,使用者可借助该工具下载与硬件规格相匹配的模型和容器,并将模型连接至 OpenCode。工具还将自动加载 DGX Spark Skills,为设备故障排查、维护及工作流程给出指导。

软件更新维度,10 月版 DGX Spark 更新加入 Wi-Fi 助手,允许设备在 appliance mode 下切换无线网络;DGX Spark Skills 也将借助 OpenCode 集成至英伟达 Sync,为日常维护和故障排查给出兼容。

模型维度,英伟达宣布兼容更多可在本地运行的 AI 模型,涵盖 Nemotron、GLM、DeepSeek、Qwen、MiniMax 以及 FLUX.3 Klein 等,覆盖智能体、编程、图像生成和编辑等任务。

英伟达还推出针对本地 AI 硬件优化的 Recursive Flint 编程模型。该模型使用混合专家(MoE)架构,总参数量为 400 亿,激活参数量为 25 亿。英伟达称,该模型在配备 64GB 内存的 RTX Spark 上推理速度超过每秒 220 个词元。

针对本地运行大型模型时的内存压力,英伟达还发布了一组优化内存使用的模型检查点,借助量化降低内存占用,同时尽可能保持接近全精度模型的准确性。其中,Qwen3.8-Flash-Next-FP4-Local 可在单台 128GB DGX Spark 或 RTX PRO 6000 上运行;GLM-5.3-Flash-FP4-Local 则面向两台平台的规格。

推理表现优化是此次更新的另一项重点。依据英伟达公布的数据,在 llama.cpp 中,针对 Qwen 3.8 Flash Next 的多 token 预测(MTP)推测解码优化,可使 RTX PRO 6000 Blackwell 上的端到端智能体任务提速最高 1.4 倍;在 vLLM 中,针对 Qwen3.8 27B 的推测解码和 FlashInfer 预填充内核优化,可使 DGX Spark 上的相关任务提速最高 1.3 倍;DGX Station 运行 GLM 5.3 Flash 时,吞吐量最高提高至原来的 1.2 倍。

英伟达还公布了 DGX Station 的企业管理与应用扩展计划,涵盖新增 Absolute Security、思科旗下 Splunk、Microsoft Intune、Tanium 等七项管理集成,协助企业部署和管理本地 AI 基础设施。

DGX Station 的 Windows 版本预计于 12 月推出,将兼容制造业相关应用,以及外观、仿真、验证和工程自动化工作流。

另外,英伟达计划为 DGX Station 推出开放源代码机器人参考工作流,并以早期访问形式给出 NVIDIA Cloud Functions,让团队借助统一接口共享本地模型服务。

英伟达 DGX Spark 支持 8 台集群 128GB 版本统一内存最高达 1TB RTX PRO 6000 Blackwell 相关任务提速 40%

英伟达 DGX Spark 支持 8 台集群 128GB 版本统一内存最高达 1TB RTX PRO 6000 Blackwell 相关任务提速 40%

相关阅读:

本文来源:https://www.ithome.com/1/016/148.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT House October 21st news, Nvidia announced today at the GTC 2026 conference in Berlin, DGX Spark was previously compatible with up to 4 devices to form a cluster, and now up to 8 devices of manual specifications have been opened to form a cluster. Up to 1TB of unified memory if built in 128GB version.

At the same time, Nvidia also released an update to the local AI platform and announced that it was working with the llama.cpp and vLLM open source communities to optimize AI inference performance. Under specific workloads, RTX Pro 6000 Blackwell's inference performance increased by up to 40%, and DGX Spark and DGX Station also achieved up to 30% and 20% performance improvement, respectively.

Nvidia said a 64GB version of the $4,999 DGX Spark will be available starting October 23 with partners. For cluster deployments, the Nvidia Sync Model Launcher is scheduled to go live later this month, allowing users to download models and containers that match the hardware specifications and connect the models to OpenCode. The tool will also automatically load DGX Spark Skills to guide equipment troubleshooting, maintenance, and workflows.

Software update dimension, October version of DGX Spark update added Wi-Fi assistant to allow devices to switch wireless networks in appliance mode; DGX Spark Skills will also be integrated into Nvidia Sync with OpenCode to provide compatibility for routine maintenance and troubleshooting.

NVIDIA DGX Spark Supports 8-Unit Clusters; 128GB Unified Memory Tops 1TB with RTX

Model dimension, Nvidia announced that it is compatible with more AI models that can be run locally, covering Nemotron, GLM, DeepSeek, Qwen, MiniMax, FLUX.3 Klein and other tasks such as agent, programming, image generation and editing.

Nvidia also introduced a Recursive Flint programming model optimized for native AI hardware. The model uses a mixed expert (MoE) architecture with 40 billion total parameters and 2.5 billion activated parameters. According to Nvidia, the model reasoned at more than 220 words per second on RTX Spark with 64GB of memory.

In response to memory pressures when running large models locally, Nvidia also released a set of model checkpoints that optimize memory usage, reducing memory footprint with quantization while maintaining accuracy as close to full-precision models as possible. Among them, Qwen3.8-Flash-Next-FP4-Local can run on a single 128GB DGX Spark or RTX Pro 6000; GLM-5.3-Flash-FP4-Local is for two platform specifications.

Reasoning performance optimization is another focus of this update. According to data released by Nvidia, in llama.cpp, speculative decoding optimization for Qwen 3.8 Flash Next Multi-Token Prediction (MTP) can speed up end-to-end agent tasks on RTX Pro 6000 Blackwell by up to 1.4 times; in vLLM, speculative decoding for Qwen 3.8 27b and FlashInfer prefilled kernel optimization can speed up related tasks on DGX Spark by up to 1.3 times; when DGX Station runs GLM 5.3 Flash, the throughput is up to 1.2 times higher.

Nvidia also announced an enterprise management and application expansion plan for DGX Station, covering seven new management integrations, including Absolute Security, Cisco's Splunk, Microsoft Intune, and Tanium, to help enterprises deploy and manage on-premises AI infrastructure.

NVIDIA DGX Spark Supports 8-Unit Clusters; 128GB Unified Memory Tops 1TB with RTX

The Windows version of DGX Station, expected to launch in December, will be compatible with manufacturing-related applications, as well as appearance, simulation, validation, and engineering automation workflows.

In addition, Nvidia plans to launch an open source robotic reference workflow for DGX Station and provide NVIDIA Cloud Functions in the form of early access, allowing the team to share local model services through a unified interface.

This article is from: https://www.ithome.com/1/016/148.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

NVIDIA DGX Spark Supports 8-Unit Clusters; 128GB Unified Memory Tops 1TB with RTXNVIDIA DGX Spark Supports 8-Unit Clusters; 128GB Unified Memory Tops 1TB with RTX

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry