Qualcomm Charges Into Data-Center CPUs
On June 25, at its Investor Day, Qualcomm laid out its data-center roadmap all at once: the Dragonfly C1000 CPU, High-Bandwidth Computing (HBC), the Dragonfly AI300 inference accelerator, plus accompanying connectivity products. The most attention-grabbing is the C1000—a chiplet design with over 250 cores, a clock speed as high as 5GHz, and support for PCIe Gen7, CXL, and LPDDR memory. The 5GHz figure itself is notable, because same-generation Intel Clearwater Forest and AMD EPYC Turin Dense don't hit such high frequencies.
Customized for Agentic AI
Qualcomm's headline selling point this time is "agentic AI." Its argument: in agent-type workloads, quite a few operations are actually better suited to CPUs than to the massively parallel architecture of piling on GPUs; and the metrics it stresses shift from raw compute to "performance per watt" and "total cost per token." Commercially, Qualcomm has already brought in Meta for an endorsement—Meta plans to adopt the C1000 after it enters mass production in 2028; and Qualcomm has set a target for its data-center business: $15 billion in revenue by 2029. Worth noting is that most of this is still roadmap and promises: the chip won't reach mass production until 2028, and the revenue target is years out. In the short term it can't shake Nvidia's position in AI compute, but the direction of "CPU makers eyeing agentic AI" itself shows that competition is spreading from training-side GPUs to the compute structure of inference and agents.
via: Qualcomm