GPU 基礎設施

Anthropic 已向美國證券交易委員會提交 S-1 招股說明書，披露了高管薪酬、算力支出計劃、特殊風險因素、公司治理結構以及它所認定的競爭對手。

本文系統講解大語言模型推理的全過程：從分詞、預填充到逐 token 解碼，KV 快取如何讓模型“記住”上文，為什麼推理會受算力、視訊記憶體頻寬、長上下文與併發請求的制約，以及量化、FlashAttention、連續批處理、PagedAttention、張量並行和投機解碼等最佳化手段如何分別解決不同瓶頸。文章還給出了衡量推理效能的關鍵指標，並說明構建自有效能測試實驗的思路。

研究人員提出 modded-nanoRT，這是首個面向關係型 Transformer 的“速通”排行榜。固定任務、開放技術棧，以單塊 GPU 達到質量目標所需實際耗時為排名標準，從而把預訓練配方探索從單個實驗室的私有協議中解放出來。團隊還引入按外部索引鍵路徑調整注意力分數和值的型別化三線性注意力，配合 Muon 與 SIGReg，在單塊 NVIDIA H200 上以 305.63 秒達到目標，比參考實現快 8.11 倍，並在開發資料庫之外的 15 組對比中贏得 13 組。

哈薩克的AI生態正從基礎設施規劃走向實際部署，覆蓋從底層算力到模型再到應用的完整技術棧。國有電信巨頭Kazakhtelecom已在阿斯塔納和阿拉木圖部署400塊NVIDIA H200與480塊B300 GPU，並計劃明年在Data Center Valley增加逾1000塊GPU、在埃基巴斯圖茲地區部署逾1000塊Vera Rubin GPU。與此同時，inDrive與Pony.ai正探討在哈薩克推出Robotaxi服務。

在Meta財報電話會議上，馬克·扎克伯格把公司的未來押注在個人AI智慧體Muse上，設想了一個由數十億個Muse相互對話和交易組成的網路，並聲稱它將比Instagram和Facebook帶來更多單使用者收入。但該計劃依賴於削減算力成本，而Meta尚未解釋將如何實現這一目標，同時隱私方面的擔憂和低迷的消費者信心仍揮之不去。

亞馬遜雲科技（AWS）宣佈 SageMaker AI 無伺服器模型定製功能全面推出訓練會話（training sessions）。使用者可自行編寫並協調後訓練迴圈，包括自定義損失函式、獎勵訊號和多輪強化學習，而 GPU 分配、分散式執行、故障恢復與檢查點儲存均由 SageMaker AI 託管。該功能支援 gpt-oss-20b、gemma-4-31B-it、Qwen3.6-27B 和 Qwen3.8-27B，按處理的 token 計費。

Amazon SageMaker AI 推出訓練會話 API，允許開發者用 Python 編寫自定義 LLM 訓練迴圈（如 SFT、RL 或自定義目標），而服務負責管理 GPU 基礎設施。該 API 提供 forward_backward()、optim_step() 和 sample() 等操作，按 token 計費，並配有包含 SFT 和 RL 示例的 cookbook。

Modal CEO Erik Bernhardsson在訪談中解釋了公司如何透過池化需求提供彈性GPU容量，實現年化收入超10億美元、完成7.5億美元D輪融資。他討論了GPU需求非矩形、沙盒產品意外增長、快速啟動作為基礎設施功能等話題。

Modal 在完成 7.5 億美元 D 輪融資、估值達 160 億美元后，CEO Erik Bernhardsson 與 Latent Space 對話，講述公司如何將 GPU 容量從“矩形”合約轉化為按需彈性算力、沙箱業務意外成為約三分之一收入來源，以及 AI 基礎設施的經濟學與工程文化。

總部位於巴黎的語音AI公司Gradium（由Kyutai分拆而來）開源了Phonon系列端側文本轉語音模型。其中英語版Phonon 31m僅有31M引數、64.5 MB大小，採用Apache 2.0許可，無需雲端呼叫、GPU或按字元計費。Gradium的評測顯示，它在難句測試集上詞錯率為1.70%，遠優於Kokoro的3.50%，並可在筆記型電腦CPU上約7毫秒內輸出首個音訊。

AWS 部落格介紹了一種以實測資料替代猜測的自動擴縮容方法：在 Amazon SageMaker AI 即時端點上以自定義容器部署 Kokoro-82M 文本轉語音模型，透過對線上端點測量 Q_max（單例項在 SLO 內可承受的最大併發）和 T_total（端到端擴容耗時），再用算術推匯出擴容閾值、縮容閾值以及兩段冷卻時間，從而避免因閾值設定偏保守造成 GPU 閒置浪費、或偏激進導致延遲 SLO 被擊穿卻不被察覺的問題。

輝達推出 AIOps 平臺 Minerva，把基礎設施與 AI 工作負載的遙測資料匯聚起來，並藉助專用 AI 模型把海量訊號轉化為可執行的洞察，幫助團隊更早發現問題、提升 GPU 利用率與 AI 工廠產出。IREN 等多家合作伙伴計劃部署，戴爾、聯想、Supermicro 將把它納入其 AI 工廠運營產品組合。

輝達GeForce NOW本週四迎來《使命召喚：現代戰爭4》首發，同步上線戰役、多人模式和DMZ撤離模式，並藉助RTX 5080級效能為終極會員帶來光追體驗。此外，Fire TV Stick使用者現可直接透過亞馬遜購買會員，史克威爾艾尼克斯的《最終幻想：共鳴》與《上古卷軸OL》新賽季等六款遊戲同步加入雲遊戲庫。

Modal 釋出 Spindle，一個開源且相容 Tinker 的訓練伺服器，可執行在使用者自己的 Modal 工作區內。它讓團隊用 GPU 時長而非按 token 計費，並能自定義模型與功能，只需一條命令即可部署。

在柏林GTC大會上，多家歐洲銀行、支付公司和金融資料提供商展示了基於NVIDIA加速計算、開放權重模型和自有基礎設施的生產級AI系統。Revolut以PRAGMA基礎模型重構技術棧，Starling Bank、Lloyds、bunq、NatWest、Santander、LSEG、HSBC和Stripe等機構也在主權雲、交易資料建模、代理助手、語音識別和欺詐檢測等場景中取得顯著效率與成本收益。

NVIDIA 透過部落格展示歐洲零售企業如何藉助開放模型與物理智慧改造業務：Shopify 後訓練 Nemotron 模型將推理成本降低 90%、響應速度提升 7 倍；Zooplus 產品發現原型實現 86% 搜尋準確率；Otto Group 在德國最大履約中心用 Omniverse 數字孿生在模擬中設計並最佳化倉庫運營；Nebius 在 GTC 柏林展示三套零售藍圖。

一篇 arXiv 論文探討了 Gemma 4 與 EmbeddingGemma 2 共用的 3.05 億引數音訊編碼器在不改動任何權重的前提下，用於沙特與海灣阿拉伯語語音識別的效果。作者構建了兩個系統：僅訓練 CTC 讀取頭的凍結方案，以及在其上增加低秩介面卡、混合 CTC/注意力讀取頭與子詞語言模型聯合解碼的增強方案。增強系統在多個公開測試集上與 20.7 億引數的 Cohere Transcribe Arabic 在詞錯誤率上統計不可區分，部分場景顯著更優，並可在單張 GPU 上達到 139 倍即時速度；但其在長篇沙特語音的詞錯誤率上仍落後，在盲測公共呼叫中心樣本中排名 15 個系統中的第 4。

本文介紹如何使用 NVIDIA 開源工具 Nsight Python 在 Python 環境中剖析 GPU 核心。透過在程式碼中標註關注區域並用裝飾器包裝基準函式，開發者可以直接獲得按註釋區域、配置和執行次陣列織的 pandas DataFrame 硬體指標，無需再依賴 Nsight Compute 命令列報告。教程在免費的 Colab T4 上演示了安裝、單核心剖析、執行次數選擇、快取與時鐘控制、矩陣尺寸掃描、硬體計數器判斷計算瓶頸、核心融合測量以及 CUPTI 後端對比等完整流程。

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：GPUs power today's most advanced AI workloads, from cutting-edge language models...

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Every enterprise company is currently trying to figure out how to have its AI tools work with its data. Meanwhile, The post Nvidia bets open models can give enterprise AI agents the context they’re missing appeared first on The New Stack.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Nvidia wants new AI capacity to show up at the data center fully built. At its GTC Berlin conference on The post Nvidia wants AI factories built in real factories appeared first on The New Stack.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：At GTC Berlin, NVIDIA is showcasing its full lineup of local AI hardware. RTX Spark, which launched last week, powers personal agents on Windows PCs; DGX Spark supports developers building and scaling applications on Linux; and DGX Station serves enterprise teams running frontier workloads on Linux, with a Windows version arriving in December. Announcements include: […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：General availability of the instances is coming next month, with planned expansion of availability to Europe — bringing an integrated platform for training, inference and large-scale reinforcement learning to cloud customers. A5X instances bring together NVIDIA Vera Rubin NVL72, NVIDIA ConnectX-9 SuperNICs, Google Cloud Virgo Network and Google Cloud’s fourth-generation advanced liquid-cooling technology for demanding […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Demand from AI-natives is pulling the European AI cloud ecosystem forward faster than ever. Developers need compute closer to their data, open models they can customize and serve at scale, and full-stack infrastructure that is secure by design. NVIDIA Cloud Partners across Europe, the Nordics, Africa, and Central Asia are expanding regional compute capacity for […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：To increase productivity, accelerate innovation and create new opportunities for economic growth, Europe’s industrial leaders are tapping NVIDIA CUDA-X and Omniverse libraries, as well as NVIDIA Cosmos and Nemotron open models, across the engineering and manufacturing lifecycle. Bringing those capabilities together on the factory floor, the NVIDIA Factory Operations Blueprint (FOX) is now available in […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：NVIDIA cuDSS, launched today in general availability, is an NVIDIA CUDA-X library that shifts a critical part of engineering and scientific simulations onto GPUs. Complex engineering feats — from advanced chips to aircraft to industrial systems —– depend on simulation to predict how designs will perform before they’re built. As designs become more complex, simulations […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：NVIDIA builds tools to enable safe agents across the stack. The NVIDIA OpenShell Reference Architecture, announced today, is part of NVIDIA’s growing portfolio of technologies that enables organizations to safely deploy agents at scale. NVIDIA Confidential Computing is the infrastructure layer that protects the models and data those agents depend on while running. Today, NVIDIA […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Agents are only as useful as what they know, and only as trustworthy as the boundaries they run within. Leading enterprises are using the open source NVIDIA Agent Toolkit — now expanded with new tools — to customize and control agents, use company data to inform their answers and actions, monitor what they do and […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：An open source “bomb” drops from China

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：An open source “bomb” drops from China

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Bring RFD4-Proteina to your research data with AI-assisted development and serverless NVIDIA GPU compute....

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Google, Red Hat and IBM released llm-d 1.0, an open-source Kubernetes-native stack for distributed LLM inference. It adds multi-tenant flow control, zero-downtime upgrades and agent-aware routing across vLLM, SGLang, TPUs and GPUs. The post Google Cloud Ships llm-d 1.0 With Cache-Aware Routing for TPU and GPU Fleets appeared first on MarkTechPost.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.27199v1 Announce Type: new Abstract: The Jetson AGX Orin is a common computer for robots and vehicles, yet measured accuracy, speed and energy of current depth estimators on it are hard to find. We benchmark nine model families as of October 2026: the monocular Depth Anything V2 and 3, Depth Pro, MoGe-2, Metric3D v2 and UniDepth V2, and the stereo methods libSGM, NVIDIA ESS and Fast-FoundationStereo. Every method runs on the Orin as a TensorRT FP16 engine or native CUDA code at up to three input sizes, up to the full 5.7 MP, and is scored for metric accuracy against Middlebury 2014 ground truth (non-occluded pixels of the 10 scenes whose disparity stays below 256 px), GPU time, energy and licence. Stereo is far more accurate. At 30 fps, NVIDIA's ESS…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.26915v1 Announce Type: new Abstract: Online action detection (OAD) identifies ongoing actions in streaming videos. Many existing OAD systems combine RGB appearance with optical flow, but dense-flow extraction remains a major computational bottleneck. Codec motion vectors (MVs) are available during decoding, avoiding a separate optical-flow computation. On THUMOS'14, optical flow exceeds MVs by 11.2 mAP points with motion-only inputs, whereas the gap narrows to 3.7 points after RGB fusion, with the largest reduction in action-category discrimination. We propose CodecMotion-OAD, which replaces optical flow with decoder-provided MVs while retaining existing temporal detector architectures. Its two-stage training first aligns MV features with a pretraine…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.26937v1 Announce Type: new Abstract: Large clinical language models often require multi-GPU infrastructure, limiting local deployment. We introduce CERTA, a decision-level statistical certificate that determines when a quantized model may substitute for its native BF16 reference and escalates decisions that cannot be certified to it. Across eleven instruction-tuned models, both general-purpose and domain-adapted, from three families (7-72B), we evaluated 2,500 clinical-verification decisions from de-identified patient contexts (2,484 patients), with patient-disjoint calibration and testing. In the higher-power operational analysis, round-to-nearest 8-bit quantization at BF16 storage served 99.8-100% of held-out decisions across all eleven models; for…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.26913v1 Announce Type: new Abstract: The adoption of Large Language Models for radiology report structuring has been largely dominated by either proprietary cloud-based models or large open models requiring high-end GPUs for deployment, which is a critical bottleneck for clinical adoption in radiology workflows that demand data privacy and offline operation. To address this, we present MedPAO-Assist, an edge-deployable agentic framework that generates structured radiology reports directly from radiologist dictation and voice commands, requiring only $\approx3.3$ GB peak RAM and achieving an average end-to-end latency of $\approx31$ seconds on a standard laptop CPU through an asynchronous, multi-threaded execution pipeline. The system currently suppor…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：From GPU trading to physical settlement, independent markets, and programmable compute

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：From GPU trading to physical settlement, independent markets, and programmable compute

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：The midterms will be a referendum on the GOP’s alignment with tech overlords over the public they’re meant to serve Welcome back to How We Win, a Guardian opinion series with leading thinkers, offering bold ideas for America’s political future. Last week, Robert Reich highlighted the link between corruption and affordability. Today’s edition was written by Guardian US columnist Rebecca Solnit. Sign up for future editions here. “The world should accept some bad things happening,” Sam Altman, the head of OpenAI, said recently, “for the benefits of this technology and people having the agency.” Like so many things tech oligarchs say, it’s incoherent, since Altman would like to deny the world and the people in it the agency to decide whether or not we should accept…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：DeepSeek has open-sourced the inference engine it uses to serve its models in production, a step it said last year The post DeepSeek open-sources the inference engine behind its API. And Huawei gets equal billing with Nvidia. appeared first on The New Stack.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：DeepSeek has open-sourced DeepInfer, the Apache 2.0 inference framework that serves DeepSeek-V4.1-Flash in production. It targets NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, and ships offline inference only. The post DeepSeek Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash appeared first on MarkTechPost.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：As SpaceX impacts Ukraine and AI-generated music divides artists and fans, critics are asking what happens when technology has too much control over our lives This was originally published in TechScape, a newsletter about how technology shapes our lives. Sign up to receive it here Hello, and welcome to TechScape. I’m Blake Montgomery, US tech editor at the Guardian, writing to you while sitting in front of Roy Lichtenstein’s monumental Drowning Girl, on view in New York City. Today in tech, we’re discussing the perils of Elon Musk’s one-man global infrastructure, AI-generated music, and a Guardian AI conference in Paris. Continue reading...

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Nebius and Nscale are among global technology partners looking to NVIDIA Vera CPUs to build the next wave of infrastructure. Their momentum reflects rapid advancements in agentic computing. Agents do more than generate answers: they retrieve information, run software tools, execute code and evaluate results. Each step depends on fast CPU processing. Vera accelerates that […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：The Barcelona Supercomputing Center (BSC)’s AI Factory Accelerator Programme, now in full operation following a pilot initiative, gives startups access to sovereign compute, technical support and training to support innovation across industries. Many of these companies are also members of the NVIDIA Inception program for cutting-edge startups — and as of today, access to some […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Discussion | Link

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：In the cavernous hall of Kraftwerk, a former power station in Berlin, German Chancellor Friedrich Merz and NVIDIA founder and CEO Jensen Huang sat down to discuss Germany’s AI roadmap, and in particular, the industrial applications of AI. Huang opened by tracing NVIDIA’s longstanding relationship with German industry, beginning with gaming systems, workstations and automotive […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：At a creative industry event that took place today in Berlin — hosted in collaboration with Adobe, Bertelsmann, NVIDIA and UFA — the latest advancements in intelligent media were on display. Key to the news: the European Broadcasting Union (EBU) is collaborating with NVIDIA to expand its sovereign framework with cloud, open models and integration […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Today, I’m talking with Illinois Governor JB Pritzker, who has a lot of thoughts about AI, tech, and privacy. If you’ve been listening to the show, you know that one of my running theories is that those issues are driving a big and maybe kind of unexpected political realignment — and with the midterms coming up next month, and early voting already underway in a lot of places, the stakes are pretty high right now. So I asked Gov. Pritzker about all that, and he was surprisingly willing to dive right into the absolute existential weeds of what government is even for. You’ll hear him say — and it’s hard to disagree — that some problems are just collective action problems that require the government to do a lot of work. We can’t solve all these problems one at a ti…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.22380v1 Announce Type: new Abstract: Continuous diffusion treats the forward process as a design space; discrete language diffusion often uses absorbing corruption, leaving context fully trusted throughout generation. This makes trust binary: corrupted context is copied as ground truth, propagating errors into generated spans. We reopen this design space by combining absorbing and uniform-mixing kernels with region-dependent rates. Missing positions follow absorbing dynamics, while context undergoes low-rate random substitution, exposing the model to observed but uncertain tokens during training. At inference, a heuristic sampler re-injects noise into suspicious context under an annealed intensity and freezes context updates late in the trajectory. A…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2610.22446v1 Announce Type: new Abstract: Neural operators approximate solution operators of partial differential equations at a fraction of the cost of numerical solvers. Efficient grid-based architectures such as the Fourier neural operator (FNO) assume uniform rectangular grids, whereas engineering simulations use large, irregular, and locally refined meshes. We introduce L-SplineFNO, a memory-efficient neural operator that couples an FNO backbone with a sparse, learnable tensor-product B-spline transfer between unstructured meshes and a latent Cartesian grid. A local normalised scatter avoids the geometry-dependent solve of analytic spline transformation while preserving compact support, partition of unity, and spline smoothness by construction. The n…