AI Software Engineer Intern

About This Role

• Intel's Data Center Network Edge AI team is responsible for delivering best-in-class AI performance on Intel® architecture.
• From hyperscale data centers powered by Intel® Xeon® processors to network edge nodes, our performance engineers shape the inner loops of frameworks and operator libraries that millions of developers and customers rely on every day.
• We are seeking an intern to join our CPU performance engineering team and drive operator-level optimizations for modern AI workloads, including Transformer-based LLMs, VLM / VLA multi-modal models, classical CNNs, and MLP models, etc.
• You will design, implement, and tune high-performance CPU kernels that translate Intel architectural advantages — AVX-512, Intel® AMX, and VNNI — into measurable end-user value.
• Responsibilities Design and hand-tune CPU kernels for Transformer operators (Attention, GEMM, LayerNorm, RMSNorm, RoPE, MoE, Softmax) and classical operators (Conv2D / Conv3D, Depthwise Conv, Winograd, im2col, Pooling, BatchNorm, RNN / LSTM / GRU).
• Develop SIMD-optimized implementations using Intel® AVX2 / AVX-512 / AMX / VNNI intrinsics, with ARM Neon / SVE as a secondary target where applicable.
• Apply parallelization strategies (OpenMP, TBB, thread-pool design) and exploit CPU micro-architectural features: cache blocking and tiling, NUMA affinity, prefetching, memory alignment, and false-sharing mitigation.
• Implement and optimize low-bit quantized kernels (INT8 / INT4 / W4A16 / W8A8) for LLM / VLM inference, leveraging Intel® AMX and VNNI for maximum throughput per watt.
• Integrate custom operators into production frameworks and runtimes, including Intel® oneDNN, PyTorch CPU backend, ONNX Runtime, llama.cpp, MLC-LLM, and XNNPACK.
• Conduct systematic performance analysis using Intel® VTune™ Profiler, Linux perf, and roofline modeling; identify bottlenecks and quantify optimization gains.
• Contribute reusable kernels, optimization templates, and best-practice documentation to Intel's internal performance libraries.