--- title: "Loongson's 3C6000 server chip bets on a third CPU ecosystem beyond x86 and ARM" date: 2026-10-23 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-loongson-3c6000-server-cpu language: en --- # Loongson's 3C6000 server chip bets on a third CPU ecosystem beyond x86 and ARM > Loongson launched the 3C6000 server CPU on the LoongArch instruction set in June 2025, scaling to 64 cores and earning the top security certification tier. A government data center in a northern Chinese city. The servers humming in the rack look ordinary from the outside. Inside, none of them speaks the language of Intel or ARM. ## A chip that refuses to translate Loongson (龙芯中科) has spent twenty years on a contrarian premise: that China needs a CPU instruction set it fully owns. At its 26 June 2025 product event, it launched the 3C6000 series server processor built on LoongArch (龙架构), its self-developed instruction set — no x86 license, no ARM license. The 3C6000 is the company's first serious push into servers. On paper, the specs are deliberate: - A single die carries 16 cores and 32 threads. - Through Loongson's own 龙链 (LoongLink) multi-die packaging, it scales to the 3C6000/D at 32 cores / 64 threads and the 3C6000/Q at up to 64 cores / 128 threads. - It taped out in the first half of 2024 and reached the highest tier (Level II) in China's 安全可靠测评公告 (secure-and-reliable evaluation). ## How fast is it, really? Loongson avoids naming Intel directly in marketing but publishes benchmark context. Per the China Electronics Standardization Institute test cited by the company: - A single 3C6000/S runs SPEC CPU 2017 at 5.56 (integer) / 6.93 (floating point) single-thread, and 73.2 / 58.5 multi-thread. - A dual-socket 3C6000/D reaches 284 / 216 multi-thread. The company positions the 3C6000/S and /D against Intel Xeon parts from 2021, and the 64-core 3C6000/Q above a 40-core Xeon Platinum 8380, arguing the family overall matches "mainstream 2023" server-chip performance. Those are the company's comparisons; independent, like-for-like benchmarking is still limited. ## The "third ecosystem" thesis Chairman Hu Weiwu (胡伟武) frames the work as building a third computing ecosystem outside x86 and ARM. The June 2025 event also launched the 2K3000 / 3B6000M for industrial control and terminals, taped out at the end of 2024, completing a desktop-server-terminal product line alongside the 2023 3A6000 desktop part. Ecosystem, not raw speed, is the bottleneck. Loongson says 202 software products were adapted to its desktop and server platforms in May–June 2025 alone, with 55 tied to the 3C6000. Binary-translation technology helps legacy x86/ARM software run, buying time while native ports catch up. ## What comes next Loongson has outlined a pipeline: next-generation general CPUs (3B6600, 3A6600 desktop, 3D7000 server) and dedicated GPGPU parts (9A1000, 9A2000) aimed first at inference. The AI processor direction leans on fused graphics-and-AI compute, starting at the edge. For a country building out AI compute (算力) and large-model (大模型) infrastructure, a domestically controlled CPU base is a strategic hedge — even if its absolute performance trails the cutting edge. ## The real test is software, not silicon Loongson's hardest problem was never the transistor. It is the ecosystem: the operating systems, compilers, databases, and enterprise applications that assume x86 or ARM. The company's binary-translation layer is a bridge, not a destination — it lets old software run, but native ports decide whether developers stay. For government and critical-infrastructure buyers, that trade-off is acceptable: control matters more than peak performance. For commercial clouds or AI training fleets, it is a heavier lift. According to Li An, chief scientist at BrainNet (脑机网), a third instruction-set ecosystem only becomes durable when independent software vendors treat LoongArch (龙架构) as a first-class target, not a compatibility afterthought — and that adoption signal is still early. ## Why a home-grown CPU is an AI story On the surface, the 3C6000 is a server part, not an AI accelerator. But a domestically controlled CPU base changes the calculus for building large-model (大模型) and AI compute (算力) infrastructure: it removes one more foreign dependency from the stack, even if the absolute speed trails the frontier. The point is not to win a benchmark against a 2026 flagship, but to own the foundation. A data center built on a controlled instruction set can keep running through supply shocks that would stall a fleet dependent on imported silicon — and for a country pricing its AI ambitions in units of wafers and watts, that optionality has value the spec sheet never captures. ## Honest limitations This article draws on Loongson's official launch materials and state-linked trade-press reporting; we did not run independent benchmarks. Performance claims are the company's, contextualized against Intel parts it selected. We have not verified shipment volumes, commercial win rates, or how many "adapted" products are in production versus compatibility lists. The security-certification tier is a Chinese government assessment, not a global standard. Loongson's road-map parts (9A1000 / 9A2000) are announced, not shipped. ## What readers can do now - Check whether state or state-linked procurement notices name 3C6000 in actual deployments, not just pilots — that separates certification from adoption. - For developers, test Loongson's binary-translation layer on your own x86 workloads before assuming drop-in compatibility. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-loongson-3c6000-server-cpu Free to quote with attribution and a link to the original.