[ByteDance's AI Chip Development Still Lacking Large-Scale Deployment After 7 Years: Performance Falls Short of Expectations After Tapeout, Head Resigns] According to an exclusive report by LatePost, Shi Yunfeng, head of ByteDance's AI chip and DPU team, has submitted his resignation and is currently handing over his responsibilities. The AI chip that ByteDance completed tape-out this year has not met performance expectations and has not yet been deployed on a large scale in its internal operations. The team is currently developing the next generation product. ByteDance established its AI chip team in 2019, choosing a dedicated ASIC (Application-Specific Integrated Circuit) approach similar to Google's TPU, hoping to optimize computing efficiency and energy consumption for its own AI business. Currently, Ada for Doubao and SeedChip for Seedance are still working on operator adaptation, that is, ensuring that various computational operations in the model can run correctly and efficiently on the self-developed chip, but progress is limited. The difficulty of this approach lies in the software ecosystem. NVIDIA GPUs already have a mature CUDA toolchain, while ByteDance's dedicated chips require additional adaptation of models, computing frameworks, and underlying software. Google has invested a large engineering team in its TPU, developing supporting tools over a long period. ByteDance's business units, however, urgently need more computing power and find it difficult to dedicate manpower to repeatedly debugging immature chips; directly purchasing mature GPUs is a more feasible way to meet their needs. ByteDance just restructured its chip business in February of this year. Wang Jian, the former head, no longer oversees chips; Shi Yunfeng's AI chip and DPU teams, and Yu Hongbin's CPU team, now report directly to Vice President Yang Zhenyuan. Now, Shi Yunfeng, who has been in charge of AI chip R&D for many years, has also resigned. Besides internal R&D, ByteDance is also trying to introduce external chip technology. ByteDance had planned to merge an invested chip project into its in-house R&D team, Picoheart, but the deal has not yet been finalized. Another approach is to collaborate with Xinyuan Semiconductor to develop AI inference chips using the new ReRAM storage technology. Xinyuan's latest prototype uses a multi-chip integrated design, attempting to integrate more computing power, but manufacturing costs and yield rates remain unresolved. Sources familiar with the matter revealed that, according to the current design, a single 12-inch wafer can only produce enough usable chip dies to manufacture approximately four computing cards. [Original Link]