- Location
- China, Beijing · China, Shanghai
- Type
- Internship
- Department
- Engineering
- Seniority
- Internship
- Education
- PhD
- Source
- Workday
Description
NVIDIA is seeking a passionate, world-class software engineer to join its Compute Developer Technology team(DevTech). Our team has over 150 engineers across Beijing, Shanghai, Shenzhen, Taipei, Seoul, and Sydney. We understand algorithms, GPU, and real-world applications. Our mission is to connect the NVIDIA platform with developers worldwide. We dive deep into customer projects to solve performance bottlenecks. We use insights from workloads to guide next-generation NVIDIA hardware and software. If you are driven by innovation and ambition, this is the team for you!
What you'll be doing:
Working directly with key application developers to understand the current and future problems they are solving. You will build and optimize core parallel algorithms and data structures to deliver the most effective solutions using GPUs, through both library development and direct contribution to applications. This includes training and inference optimization for large language models (LLM), contributing to frameworks and open-source projects in the large language models ecosystem, such as Megatron and TRTLLM, SGLang, vLLM...
Collaborating closely with the architecture, research, libraries, tools, and system software teams at NVIDIA to influence the build of next-generation architectures, software platforms, and programming models. This includes investigating impact on application performance and developer efficiency, and turning real-world developer feedback into actionable platform improvements.
Engaging in deep optimization of high-performance operators, involving but not limited to GPU kernel optimization, instruction-level tuning, and compiler optimization. These optimizations will directly support customers or be coordinated within computation libraries and open-source projects across the community, like cuDNN, cuBLAS, and CUTLASS and Open- source libs like DeepGEMM, FlashMLA, FlashAttention, Flashinfer...
Improving communication for broad distributed large language models workloads. You will spearhead advancements in distributed training and inference by refining communication libraries(NCCL，NCCL GIN , NVSHMEM) and engaging in open-source communication libraries(like DeepEP, NCCL EP). This demands in-depth study of interconnect topologies(NVLINK) and network protocols(InfiniBand/RoCE) to design efficient data transfer strategies and methods for compute-communication overlap.
What we need to see:
A degree or equivalent experience from a university in an engineering or computer science related field. A masters or doctoral candidate is preferred.
Solid understanding of C, C++, Python, or Fortran.
Strong knowledge of software development, programming techniques, and algorithms.
Strong mathematical fundamentals, including linear algebra and numerical methods.
Background in parallel programming and accelerated computing, with comprehensive knowledge of parallel architectures and methods for performance analysis and tuning. Experience in GPU programming is desirable.
Experience in full-stack performance analysis and optimization within at least one of these areas: large language models and high-performance computing. Having expertise ranging from operator-level through framework-level to algorithm-level optimization is strongly preferred.
Experience in distributed communication optimization is highly advantageous. This involves familiarity with remote direct memory access, GPU interconnects, collective communication algorithms, and associated open-source libraries used in large-scale model training and inference.
Solid software engineering fundamentals and system architecture thinking, with the ability to build modules and drive engineering practices in complex systems.
Strong communication and cooperation abilities, with the capability to work efficiently alongside architecture, research, and software product teams to promote optimization from concept to production.
