NVIDIA says AI agents require faster individual CPU cores, arguing that the industry’s focus on higher core counts and chiplet designs has reduced single-threaded performance.
Its Vera team told Wccftech in an interview that the shift had come at the expense of instructions per cycle, or IPC, and left less chip area for features that improve per-core performance.
NVIDIA’s Vera CPU is intended to combine faster individual cores with high data-center throughput. It contains 88 custom Olympus CPU cores on a single, monolithic compute die and a custom LPDDR5X memory subsystem.
NVIDIA says Vera’s second-generation Scalable Coherency Fabric delivers up to 3.4 terabytes per second of bisection bandwidth, three times that of leading x86 chips. Keeping the cores on one compute die avoids communication between separate chiplets.
The company’s argument centers on the sequence of work AI agents perform. An agent may execute code, call external tools, retrieve information, analyze results and decide what to do next. Some tasks can run in parallel, but others must finish before the next reasoning or execution step begins.
NVIDIA argues that strong single-threaded performance and low latency allow agents to complete those sequential steps faster. It says agentic AI requires a better balance between core count and per-core speed so CPUs can keep pace with GPUs.
A CPU that takes too long to prepare a task or process an agent’s output can leave a GPU idle while it waits for instructions. Adding CPU cores does not necessarily resolve that bottleneck if individual tasks remain slow.
Chipmakers adopted chiplets as cloud data centers expanded and demand grew for processors that could handle more workloads simultaneously. Splitting a processor across several smaller dies, rather than building it on one larger die, helped them increase core counts.
NVIDIA’s Vera team says that moving data between dies can introduce communication delays and complicate memory access, potentially limiting single-threaded performance. Its criticism concerns the emphasis on core counts rather than a claim that chiplets are inherently unsuitable.
The company says it remains open to chiplet designs when they make sense. It also does not suggest that GPUs will become less important as CPU requirements change.
NVIDIA’s Grace CPU primarily acts as a host for GPU-powered systems. By contrast, the company says agentic AI is creating demand for standalone CPU racks to handle tool calls, code execution and reinforcement learning.
