LPU Chip Architecture Engineer

Responsibilities:

• Lead the overall architecture definition of LPU chips based on the static dataflow architecture, including computing array design and on-chip SRAM storage hierarchy planning. Solve core pain points of high latency and frequent data movement in large model inference scenarios.
• Collaborate with the compiler team in the early stage to define hardware microarchitecture and implement hardware-software co-design. Align scheduling logic in advance to avoid industry pain points caused by non-modifiable hardware scheduling after tape-out.
• Conduct modeling and analysis on the computing power, bandwidth and power consumption of LPU chips. Carry out architecture performance benchmarking and optimization by comparing with traditional GPU and NPU architectures.
• Track the inference requirements of MoE large models and multimodal large models, iterate and upgrade the LPU hardware architecture to adapt to next-generation large model inference scenarios.
• Participate in chip front-end design and FPGA prototype verification, cooperate with the back-end team to complete chip tape-out, and follow up chip bring-up testing and performance optimization.
• Research dataflow chip architectures of overseas benchmark manufacturers including Groq, Etched and Cerebras, and deliver competitive analysis reports and architecture iteration solutions.

Qualifications:

• Master's degree or above in Microelectronics, Integrated Circuit, Computer Architecture or related majors, with no less than 2 years of working experience in AI chip architecture design.
• Familiar with static dataflow architecture and systolic array architecture; understand the principles of Prefill and Decode dual-stage inference for large models. Prior experience in video memory and on-chip SRAM scheduling is preferred.
• Hands-on experience in AI chip / NPU / GPU architecture design, familiar with chip front-end design flow, with solid hardware-software co-design capabilities.
• Proficient in chip performance evaluation methodologies, capable of independently completing simulation and analysis of computing power, latency and power consumption.

Preferred Skills:

• R&D experience in dataflow chips or LPU chips.
• Experience in hardware adaptation for large model inference chips.
• Familiar with the fundamentals of AI compilers.
• In-depth research experience on Groq architecture.