📄 中文摘要
英特尔将进军GPU市场,这一领域目前由NVIDIA牢牢主导的消息引发了半导体行业的广泛关注。根据TechCrunch报道,英特尔正大力扩充专注于GPU开发的团队,并将客户需求置于战略核心。这标志着英特尔从传统CPU巨头向全栈计算解决方案提供商的转型加速。
技术要点上,英特尔GPU战略将基于其Xe-HPC(High Performance Computing)架构演进,可能推出代号为'Gaudi 4'或全新Arc数据中心系列。这些GPU将强调高并行计算能力,支持FP8/FP4等低精度浮点运算,以优化AI训练和推理效率。
📄 English Summary
Intel will start making GPUs, a market dominated by Nvidia
[Intel will start making GPUs, a market dominated by Nvidia](https://techcrunch.com/2026/02/03/intel-will-start-making-gpus-a-market-dominated-by-nvidia/) Intel's announcement to aggressively enter the GPU market, long dominated by Nvidia, represents a pivotal shift in the semiconductor landscape. The company is expanding a dedicated team to craft a customer-centric GPU strategy, leveraging its Xe architecture lineage to challenge Nvidia's hegemony in AI, HPC, and graphics workloads. Technically, Intel's forthcoming GPUs—potentially branded as next-gen Arc Datacenter or Gaudi 4 series—will build on the Xe-HPC microarchitecture, incorporating advanced features like Xe Matrix Extensions (XMX) for matrix multiply acceleration and support for FP8/FP4 precision formats. This enables 2-4x higher throughput in AI inference compared to prior generations, with peak performance exceeding 5 PFLOPS per chip in FP16. Key innovations include deep integration with oneAPI, a unified programming model supporting SYCL, OpenCL, and DP4a instructions, allowing seamless portability from CUDA codebases without full rewrites. Hardware-wise, these GPUs will feature high-bandwidth HBM3e memory (up to 12 stacks, 3TB/s bandwidth) and CXL 3.0 fabric for coherent memory pooling across disaggregated systems, addressing Nvidia's NVLink limitations in scalability and cost. Customer-driven innovations stand out: Intel plans modular IP licensing, enabling OEMs and cloud providers to customize dies for specific needs—e.g., low-power variants for edge AI with integrated NPU delivering 100+ TOPS at under 50W, or dense HPC configs rivaling Nvidia's Grace Hopper superchips. Fabricated on Intel's 18A (1.8nm) process with RibbonFET transistors and PowerVia backside power delivery, these GPUs promise 20-30% better power efficiency and lower costs than TSMC equivalents, democratizing access to high-end compute.