Overview In this role you will architect and optimize high-performance ML silicon for Baidu's Kunlun AI SoC, enabling distributed training across thousands of accelerators. You will collaborate across teams to balance performance, power, and latency while shaping the AI hardware roadmap. You will stay ahead of AI trends and drive partnerships to access cutting-edge technologies, translating research into scalable systems. This position offers a chance to influence AI hardware at scale in a world-class environment.
Responsibilities- Architect and create high-performance ML silicon and interconnect thousands of Kunlun accelerators for distributed training
- Develop differentiated architectural innovations for the Kunlun AI SoC roadmap
- Architect, simulate, and design ML solutions for AI products
- Develop system-level ML architectures focusing on performance, power, and latency
- Monitor AI trends and intersect them with roadmaps
- Drive partnerships for access to advanced AI technologies
- Evaluate power, performance, and cost of architectures and subsystems
- Build scalable modeling and performance evaluation tools
- Engage with system/software engineers to optimize hardware/software stack
- Collaborate with SoC design, verification, and validation teams to realize architecture
Key requirements- 5+ years of hardware solutions experience for ML, acceleration and optimization
- Experience with deep learning frameworks (TensorFlow, PyTorch, PaddlePaddle)
- Proven outreach to ML researchers and application developers
- Experience with CPUs, GPUs, memory systems, and accelerators
- Experience with performance modeling/simulation in C++
- Experience with SoC interconnects and NoCs
- Experience with area, frequency, and power optimizations
- Familiarity with video, DSP, Ethernet, and PCIe
- MS or PhD in Electrical or Computer Engineering
- Excellent communication in English and Chinese
- strong communication and collaboration
- self-directed and proactive learning
- team-oriented and cross-functional collaboration
- TensorFlow
- PyTorch
- PaddlePaddle
