EE Times · 3h ago
EE Times · 3小时前
OpenAI presented first benchmarks for Jalapeño at Hot Chips 2026, describing it as a clean-sheet general-purpose ASIC built specifically for AI workloads. The disclosure makes Jalapeño a more concrete entrant in the custom-silicon push, where performance must be judged alongside software maturity and memory-system design.
OpenAI 在 Hot Chips 2026 首次展示 Jalapeño 的基准测试，并将其定义为面向 AI 工作负载从零设计的通用 ASIC。这使 Jalapeño 成为定制芯片浪潮中更具体的参与者；其竞争力仍须结合软件成熟度和存储系统设计来判断。
SemiWiki · 4h ago
SemiWiki · 4小时前
Intel’s Crescent Island is a discrete data-center GPU aimed at agentic AI inference, where repeated model calls, tool use, weights and long-context KV caches all contribute to latency. The framing shifts attention from peak matrix throughput toward memory capacity and end-to-end response time for multi-step workloads.
Intel 的 Crescent Island 是面向智能体 AI 推理的离散数据中心 GPU；在这类负载中，反复模型调用、工具使用、模型权重和长上下文 KV cache 都会叠加延迟。这一定位将焦点从峰值矩阵吞吐转向内存容量与多步骤工作负载的端到端响应时间。
Wccftech · 1h ago
Wccftech · 1小时前
Nvidia has reportedly paused revenue-sharing and credit-support arrangements for cloud providers after internal concern about antitrust exposure. The report follows Nvidia’s disclosure of $108 billion in commitments to OpenAI and other infrastructure partners, putting the commercial structure around AI capacity under closer scrutiny.
据报道，Nvidia 因内部对反垄断风险的担忧，暂停了与云服务商的收入分成和信用支持安排。此事发生在公司披露已向 OpenAI 等基础设施合作伙伴作出 1,080 亿美元承诺之后，使 AI 产能背后的商业安排受到更严格审视。
EE Times · 7h ago
EE Times · 7小时前
Qualcomm is backing Modular’s effort to separate AI software from the underlying silicon platform. A more portable stack would give Qualcomm and other suppliers a clearer route into workloads now concentrated around Nvidia’s ecosystem, although compatibility alone does not settle performance or deployment risk.
Qualcomm 正支持 Modular 将 AI 软件与底层硅平台解耦的努力。更具可移植性的技术栈将为 Qualcomm 及其他供应商进入当前集中于 Nvidia 生态的工作负载提供路径，但兼容性本身并不能解决性能或部署风险。