At the 2026 Yunqi Conference, held in Hangzhou from September 22 to 24, Gao Hui, Vice President of T-Head Semiconductor, delivered a keynote speech. Gao noted that AI workloads in agent applications have undergone three irreversible shifts.
First, the computing focus is transitioning from training-centric to inference-first. Second, workload patterns are moving from single-round requests to multi-step orchestration. Third, the competitive metric is shifting from individual chip performance to system-level collaboration.
Gao predicts that in the agent era, a single CPU will no longer handle all scenarios. Instead, CPUs will split into two product lines, each tailored to fundamentally different workload demands.
The first type is latency-sensitive. Serving AI nodes, it handles request preprocessing and GPU kernel scheduling. Every step is intertwined with GPU inference cycles, where CPU and GPU wait on each other, and every microsecond impacts user experience. This requires strong single-core performance, high memory bandwidth, and a short interconnect path to the GPU.
The second type is throughput-concurrent CPUs. Serving agents, they handle tool invocation, retrieval, and sandbox execution. Rather than a single agent running, hundreds or thousands of agents operate concurrently, demanding many cores, large memory capacity, strong isolation, and lower per-core cost. "One pursues ultra-low latency, the other pursues massive concurrency. A single chip can hardly achieve both optimally, so we chose to design them separately, each pushed to its extreme," Gao stated.
Gao emphasized that AI's transformation goes beyond stronger models and faster chips; it is redefining the entire computing architecture. In this new system, XuanTie chips generate intelligence, Yitian CPUs drive intelligent execution of actions, networking scales computational power, and storage ensures the continuous flow of context and knowledge. In the future, AI will become ubiquitous infrastructure, much like electricity and the internet.
Comments