gpu

Ujval Kapasi of NVIDIA delivers a PyTorchCon 2026 keynote on what agentic workloads demand from the PyTorch stack, using NVIDIA Vera Rubin as a case study for hardware-software co-design across CPUs, GPUs, memory, and networking, covering optimizations in PyTorch, Triton, and vLLM.

swyx points out DeepSeek shared more substantive details on their Zhihu page, including claims that their in-house inference framework DeepInfer achieves ~155 trillion tokens/day per cluster and $38 per GPU-hour revenue, implying an ~80% serving margin versus typical $5-10/GPU-hour costs.

DeepSeek open-sources DeepInfer, the high-performance inference framework powering its large-scale production services, featuring optimized kernels, MoE expert parallelism, speculative decoding, batch-invariant inference, and support for NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

Marvell unveiled a photonic module that lets HBM and DDR memory sit up to 50 meters away from a GPU while keeping latency in the hundreds of nanoseconds, decoupling memory and compute scaling in AI clusters.

HardenedLinux 发布开源编译器优先的 AI 推理运行时 SonicBoom v0.0.3，声称在 RTX 3050 上解码速度超过 llama.cpp，并呼吁社区贡献 GPU 基准测试结果。

Three new tools let users modify and load custom VBIOS images on NVIDIA RTX 30/40/50-series GPUs on Windows and Linux without flashing, enabling higher power limits, unlocked voltage, and larger overclocking offsets while keeping the card's original firmware intact.

NVIDIA 官方證實 GeForce RTX 5090 並未停產，將繼續出貨，以澄清先前的停產傳聞。

Intel released LLM-Scaler-vLLM PV 1.5, a major update to its containerized vLLM stack for Arc Graphics hardware, jumping from vLLM 0.14 to 0.26, adding new Qwen and Gemma model support, FP8/INT4 quantization, and performance improvements.

Artificial Analysis 的 AA-AgentPerf 基准通过回放真实编码 agent 轨迹测试推理硬件能效，结果显示 NVIDIA Vera Rubin 在几乎所有速度档位上均居前沿首位，领先于 DeepSeek V4 Flash/Pro、Kimi K3 和 GLM-5.2/5.3 等模型的测试配置。

A new RTX Spark laptop (HP OmniBook X) configured with 128GB of RAM is on sale at HP for $3832.49 after a 25% coupon, offering a cheaper alternative to Nvidia's DGX Spark for under $4000.

用户提到6月购买的GPU和DGX Spark价格至少上涨了1.5倍，并指出本地AI加速的性能数据很明确。

Netflix 技术博客文章，介绍其如何在大规模训练与推理任务中检测故障 GPU 和网络、自动修复故障机器，并让长时间运行的任务在硬件故障时继续运行。文章强调了 GPU 可靠性对可用性和成本效率的重要性，并解释了为何基础健康检查不足以应对现代 ML 基础设施的需求。

tqLLM 是一个 C++/CUDA 本地 LLM 推理引擎，支持 OpenAI 兼容 API 与 TurboQuant GGUF 权重，在 RTX 3090 和 RTX 5060 Ti 上实现 240+ tok/s 的解码速度、200K+ 长上下文以及不改变输出的投机解码。

回顾Nvidia于1999年10月11日发布GeForce 256——首款被其称为GPU的产品，以及Sega当年500万美元投资如何帮助Nvidia渡过难关并最终击败3dfx的历史故事。

We trained a model on agentic coding tasks using fully asynchronous RL with mixed-precision, mixed-accelerator samplers, achieving training stability on par with homogeneous clusters at lower cost by applying off-policy correction.

Nvidia's DLSS 5 will debut in Hogwarts Legacy on November 5th with significantly toned-down facial and environmental alterations after gamer backlash, alongside an October update making the model twice as fast on RTX 50-series GPUs with 24% fewer parameters.

Intel 的 "Crescent Island" GPU 获得了 D 系列命名，并支持最高 480 GB 的 LPDDR5X 内存，主要面向数据中心 AI 加速场景。

Intel 发布基于 Xe3P 架构、专为 agentic AI 优化的 GPU D 系列，这款风冷 PCIe 5.0 显卡支持 vLLM 与开源 PyTorch 生态中的主流开放模型。

Intel 发布面向推理优化的数据中心 GPU D270，配备最高 480GB LPDDR5X 内存、32 个 Xe3P 核心与 256 个 XMX 引擎，采用 350W 风冷 PCIe 5.0 标准形态，专为 agentic AI 和长上下文推理工作负载设计。

A Twitter exchange speculating whether GPUs will eventually rent more GPUs to train improved versions of themselves, while noting that insufficient compute and token generation will remain the bottleneck for AI progress for years.