CUDA
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 9
- First Nvidia RTX Spark laptops go on sale: hands-on with local 27B inference
The first Windows laptops with Nvidia's RTX Spark superchip are on sale, and ifanr tested the 16-inch ASUS ProArt 16 RTX Spark (RTX Spark N1X), priced at 28,999 yuan for 64GB of unified memory and 37,999 yuan for 128GB. It marks Nvidia's first complete Windows compute platform — its own Arm CPU, a Blackwell RTX GPU and unified memory — rather than just supplying graphics cards; the 64GB unit exposes roughly 46.7GB of usable GPU memory with LPDDR5X bandwidth peaking near a theoretical 300GB/s. In testing, a Qwen 3.8 27B model in LM Studio generated about 14.3 tokens/s with a 2.83s time to first token, and handled long contexts without obvious degradation; under sustained local inference the machine drew about 86.5W and its 99.9Wh battery would theoretically last roughly an hour and 15 minutes at ~80W average. WorkBuddy offers a beta for downloading and managing local models on the device; the model's file header indicates a Tencent-optimized 50B-parameter MoE, 16GB in size and CUDA 13 accelerated, and LM Studio's local API can feed other tools.
爱范儿 · 🔥 6 - Microsoft Maia 300 vs Nvidia Rubin: specs compared, no clear winner
Microsoft unveiled Maia 300 on Oct. 12, 2026, and plans to expand its Azure Maia infrastructure to multi-gigawatt deployments starting in 2027, positioning the chip as its primary AI inference engine. Maia 300 pairs 288 GB of HBM4 and 18 TB/s peak bandwidth per accelerator at 950W, with 96 accelerators per rack totaling 27.6 TB of HBM4 and 167 kW; Nvidia's Rubin NVL72 lists 288 GB HBM4 and up to 22 TB/s per GPU, with 72 GPUs and 20.7 TB aggregate HBM4 per rack. Their efficiency claims use different baselines and no independent, workload-matched test exists: Microsoft says Maia 300 delivers 40% more tokens per megawatt than unnamed leading GPU systems, while Nvidia projects Rubin NVL72 at up to 10x the tokens per megawatt of GB200 NVL72, so neither can be declared faster.
TechRepublic · 🔥 4 - Microsoft details Azure Maia 300, targeting large-scale deployment in 2027
Microsoft chairman and CEO Satya Nadella announced on the evening of Oct 12 that the company plans to deploy Azure Maia 300 at scale in 2027, according to a TMTPost analysis. Microsoft says the chip uses a software-defined dataflow architecture and a chiplet design with its compute die on TSMC's N2P process, offering 18TB/s peak memory bandwidth per accelerator at 950W thermal design power, while a rack-level system of 96 accelerators provides 27.6TB of HBM capacity and 1.7PB/s of aggregate bandwidth; the company claims 40% more AI tokens per megawatt than leading GPU systems. The Maia line has moved from Maia 100 (November 2023) to Maia 200, released in January and now in production carrying Copilot workloads, to Maia 300 for 2027, though the article notes Microsoft still relies heavily on Nvidia GPUs and CUDA and says its custom silicon will not replace third-party supply.
钛媒体 · 🔥 0 - NVIDIA shows how AI agents rewrite Spark UDFs for GPU speedups with cuDF skills
NVIDIA published a technical post introducing cuDF skills for Apache Spark, a set of agent-oriented skills that guide coding agents through rewriting Apache Spark user-defined functions (UDFs) into GPU-executable implementations. The post explains that UDFs break Spark's declarative abstraction, forcing otherwise GPU-accelerated pipelines to fall back to row-by-row CPU execution with data-transfer overhead, and outlines three rewrite routes: Spark SQL expressions, the RapidsUDF interface using cuDF's Java API, or custom CUDA kernels via JNI. The skills split the work into four verification-driven loops — scaffolding, conversion, benchmarking and optimization — and ship with a project template, test harnesses and benchmark scaffolding. NVIDIA says a major consumer technology company converted more than 30 production UDFs this way; in one example a UDF used by 3,000+ jobs dropped from 274.55 to 12.78 hours of operator time on a 600-billion-row workload, a roughly 22x speedup, and states rewrites can exceed 20x. The skills are available on GitHub and installable via npx skills add NVIDIA/cudf-spark.
NVIDIA Technical Blog · 🔥 0 - Chinese tech firms bought over 1M Nvidia gaming, workstation GPUs in a year, FT reports
Chinese tech companies have bought more than 1 million Nvidia gaming and workstation GPUs this year to fill gaps in AI compute demand, according to a Financial Times report citing industry insiders. The chips are said to be mostly export-control-compliant models such as the RTX PRO 6000D, RTX PRO 5000 and RTX 5090, which lag Huawei's Ascend 910C and other domestic flagships on advanced AI work but remain useful for simpler workloads such as prefill. The report says China's domestic AI chip output roughly doubled this year to about 3.5 million units against demand well above 5 million, leaving a shortfall of nearly 1.5 million, and that Tencent has agreed to rent 100,000 GPUs in Southeast Asian data centers from Oracle, with ByteDance and Alibaba pursuing similar arrangements.
Techmeme · 🔥 0 - Qualcomm bets on open software, acquisitions to loosen Nvidia's CUDA grip
TechNews reported on Oct. 13 that Qualcomm is pushing an "open software, proprietary silicon" strategy, keeping the hardware layer closed and differentiated while making upper-layer software tools portable across platforms to erode the developer lock-in Nvidia built around CUDA. To fill in its software stack, Qualcomm has disclosed acquisitions of Edge Impulse, Arduino and Modular since early 2025, with robotics software maker PickNik Robotics added in its latest move. At Snapdragon Summit the company said it will unify Mojo, MAX and Modular Cloud into a single software layer for Snapdragon phones, X2 PCs, AR1 glasses, Wear Elite watches, Sound Gen earbuds and the upcoming Dragonfly data-center processor, letting developers move work between Nvidia or cloud GPUs and Snapdragon devices via a Mojo workflow without rewriting code. The report says Nvidia's edge lies in its integrated toolchain, simulation and robotics compute platform plus a large developer community, which also raises switching costs.
科技新报 · 🔥 0 - Iluvatar CoreX unveils TianGai 300 IXL144 rack with 144 GPUs
At the 2026 China International Industry Fair, Iluvatar CoreX launched what it calls the industry's highest-density single-rack supernode, the TianGai 300 IXL144, packing 144 GPUs and 18 CPUs into one cabinet. The system is compatible with CUDA and other AI ecosystems so customers can reuse existing models and frameworks to cut migration and deployment costs, and uses high-bandwidth whole-rack interconnect to form a unified memory-semantics compute domain for more efficient large-model training and inference. Iluvatar CoreX estimates the cost per million effective tokens is 40% lower than mainstream international solutions, a vendor figure.
36氪快讯 · 🔥 0 - Microsoft 128GB Surface Laptop Ultra sells out in US as CUDA lifts AI PC demand
Microsoft's US store shows the Surface Laptop Ultra with 128GB unified memory and a 1TB SSD is out of stock; that configuration is priced at $5,899.99, while lower-spec models remain available, and Microsoft has not said when it will be restocked. Chinese tech outlets attribute the sell-out to the machine's support for the CUDA ecosystem, since most AI frameworks, inference engines and creative tools are built for CUDA first, making it attractive to developers and creators who want to run models locally. Tech site Wccftech also suggests early supply may simply have been limited, a claim that cannot be verified because Microsoft has not released detailed figures; rival AMD Ryzen AI Max 400 systems offer up to 192GB of unified memory.
cnBeta · 🔥 0 - Lenovo YOGA Pro 15 RTX Spark opens China presales as first AI PC on NVIDIA N1X chip
Lenovo began China-wide presales of the YOGA Pro 15 RTX Spark on Oct 8, with sales starting Oct 16; the overseas version is the Yoga Pro 9n. Lenovo says it is among the first AI PCs powered by NVIDIA's RTX Spark N1X superchip, which pairs up to a 20-core Grace CPU with a 6,144-core Blackwell RTX GPU for up to 1 PFLOPS of peak FP4 local AI compute, with up to 128GB unified memory and 2TB storage, claimed local support for 100B-plus-parameter models and a preinstalled 35B local model. It ships with Lenovo's Tianxi AI agent in a new "Super Mode" that the company pitches as offline-capable, token-free and keeping data on-device.
量子位（原生 RSS） · 🔥 0
Experience and discussion from the community
Share my CUDA experienceAsk about CUDA
Nobody has shared their experience with CUDA yet.