Tag
NVIDIA 发布 RTX Spark 笔记本,将 Grace CPU 与 Blackwell GPU 封装在同一芯片上,共享最高 128GB 统一内存,AI 算力达 1 petaflop FP4,面向本地大模型运行场景。
A Reddit user reports that Dell is currently selling RTX 6000 Pro Blackwell workstation GPUs bundled in Dell desktops for around $12k, and with discounts and rewards the net cost of just the GPU could drop to roughly $10k — an unusually low price amid high GPU demand.
ISTA-DASLab 发布了 Qwen3.8-27B 的 NVFP4 量化版本,通过从 SSD 流式读取的 prefiller 模型在低 VRAM 下恢复低比特量化精度,且在超过约 4k 的上下文长度下无明显速度损失,需要 Nvidia Blackwell GPU。
作者发布了一个免费、代码优先的 GPU 内核工程系列课程(共 44 节)的介绍与路线图视频,从朴素 GEMM 讲到 FlashAttention,覆盖 Hopper 与 Blackwell 架构,并以 roofline 模型和数据移动为核心教学理念。
Basalt inference tool adds native NVFP4 support and faster prefill for Qwen3.8 Flash-Next on a single RTX 5090, delivering up to 479 tok/s decode and 8,582 tok/s prefill with near-Q8 quality retention.
NVIDIA RTX Spark laptops are now available worldwide, featuring Blackwell RTX GPUs, Grace CPUs, up to 1 petaflop of AI compute and 128GB unified memory, starting at $2,399 from major PC makers. NVIDIA celebrated with global launch events, including a Berlin appearance by CEO Jensen Huang during the Festival of Lights.
A community member benchmarks Ninfer6000 (a Ninfer fork tuned for RTX Pro 6000 Blackwell and Qwen3.8-Flash-Next) against vLLM and llama.cpp, finding ~1.6x decode speedup over vLLM and 2.2x over llama.cpp with near-identical accuracy, but reports crippling real-world tool-calling malformations that keep them on vLLM.
微软与英伟达联合发布首批 RTX Spark 笔记本,并大幅优化 Windows 以支持 Arm 架构、Blackwell GPU 与本地 AI 智能体,用户可通过 Windows Store 两步安装 Hermes、Perplexity、OpenClaw 等智能体,同时借助 Prism 模拟器改进 x64 应用兼容性与游戏性能。
NVIDIA 宣布多项英国生态合作与投资,包括五年2.5亿美元的 agentic 安全研究投入、与 Earthshot Prize 的合作,以及 Fuse Energy、Locai Labs、OneAdvanced、Speechmatics、Danaher、利物浦城市等企业与机构在能源、医疗、智慧城市等领域的全面部署。
Basalt, a Blackwell-tuned inference engine forked from Strata for Qwen 3.8 Flash Next, introduces dynamic vision encoding that keeps the image encoder in RAM when idle to save VRAM, uploading it only when an image arrives with near-zero text-inference overhead. The team also shared commit hashes so developers can port the technique to their own engines.
Decapybara 是 MegaCapybara LLM 推理服务器的开源 C++20 与 CUDA 重构实现,面向 NVIDIA RTX 5090 Blackwell 硬件,支持 OpenAI 与 Anthropic 兼容 API、分层 KV 缓存及一整套 Blackwell CUDA 内核。
Basalt 是一个专为 Blackwell 架构(5090 + 5060 Ti)深度优化的 Strata 分支推理引擎,据称吞吐量可达 Strata 的 2.6 倍,支持 Qwen3.8 Flash-Next、MTP、最多 8 路并发以及自定义视觉编码器,以 MIT 协议完全开源。
微软发布 Surface Laptop Ultra,搭载 NVIDIA RTX Spark Superchip 和 Blackwell RTX 显卡,最高配备 128GB 统一内存,可在本地运行高达 120B 参数的 AI 模型,面向开发者、创作者和 AI 爱好者。
A tweet highlights a free ~5% tokens-per-second performance boost on Blackwell GPUs by enabling CustomAllReduce on SM120 when using tensor parallelism 4, joking that closed labs hate this trick.
A hand-written GEMM kernel for NVIDIA Blackwell (B200), built with CUTLASS's CuTeDSL from raw primitives (tcgen05, TMA, mbarriers, TMEM), goes from 164 to 1401 TFLOP/s — about 97% of cuBLAS — with one key change doing most of the work and prefill vs decode workloads yielding different optimizations.
A user benchmarks the new Ninfer inference backend on a Blackwell MaxQ GPU, reporting 130-180 t/s on qwen3.8-27b-nvfp4 with minimal slowdown under parallel decoding — claiming vLLM-level concurrency and far faster throughput than llama-server, even after installing it on Windows 10.
OpenAI's GPT-6 Astra Ultrafast mode, running on NVIDIA Blackwell GPUs, is now available via the OpenAI API and to eligible ChatGPT Work/Codex users, delivering up to 8x faster token generation than Astra Standard mode. OpenAI reports using its own models to optimize inference kernels on NVIDIA hardware, improving latency, throughput, and cost for agentic coding workflows.
Release of a quantized version of the Qwen3.8-27B-DFlash2 model optimized for NVIDIA Blackwell GPUs, achieving similar decode speed and output quality with 2.6x less VRAM usage and extended context length.
This paper introduces a contract-grade verifier of twelve adversarial gates for checking LLM-generated GPU kernels, finding that 39.5% of kernels accepted by standard loose tests are broken. It also presents the first native Blackwell training backward kernel for the GDN (gated-linear-recurrence) family.
Firebird launched the CIS region's largest AI factory in Armenia, powered by NVIDIA accelerated computing and Dell infrastructure, with plans to deploy over 70,000 NVIDIA GPUs and 300 MW of capacity by 2027. NVIDIA also intends to invest in Firebird as part of a broader 2-gigawatt roadmap across Armenia, Kazakhstan, and other markets.