Chinese company Zhaoxin has announced the KH-50000 series, its most advanced server processors to date, built around a chiplet layout that clusters 12 compute dies with a large central I/O die. The top models offer up to 96 high-performance cores per socket and are available for dual- and quad-socket configurations, enabling systems with up to 384 cores. Zhaoxin quotes a base clock ranging from 2.2/2.6 GHz with boost frequencies reaching 3.0 GHz, and the family includes a substantial shared L3 cache measured up to 384 MB. Memory support expands to 12 channels of DDR5 ECC RDIMM, with per-socket capacities designed to accommodate 3 TB. The company envisions the KH-50000 as a versatile engine for cloud, AI inference, big-data analytics, and general-purpose enterprise workloads while emphasizing domestic supply chain independence and platform compatibility with existing x86 software stacks. It also includes extended platform validation and vendor support.

I/O and system interconnects form a core part of the KH-50000 story: the platform delivers 128 lanes of PCIe 5.0 with ZPI/CXL semantics, plus 16 additional PCIe 4.0 lanes and legacy SATA and USB links. Zhaoxin shows a new ZPI 5.0 interconnect designed to enhance bandwidth and reduce latency and power consumption across multi-socket configurations, enabling servers to scale performance without incurring linear energy costs. The design retains x86 32/64-bit instruction compatibility with AVX2 support and incorporates cryptographic accelerators for regional standards, aiming to resemble modern high-density server layouts. Compared to the KH-40000 family launches in 2022, Zhaoxin triples the maximum core count, transitions from DDR4 to DDR5 memory, increases cache capacity, and upgrades platform I/O to the latest generation. The firm says partners will begin validating systems soon, and it hinted at future KX consumer and AI-oriented parts that integrate NPUs for accelerated workloads. Rollout to domestic vendors is expected later in the next few months.