Loongson 3A6000: A Star Among Chinese CPUs

Original Article ↗ Hacker News Discussion ↗

Performance and Technical Characteristics

  • 3A6000 is described as Loongson’s best CPU so far, roughly between AMD Zen (14–12 nm) and Zen 2 (7 nm) per core, but limited to 4 cores.
  • Some argue that for most everyday workloads, being in the right order of magnitude of performance matters more than 10–30% deltas.
  • Others stress that “only 4 cores” is a major limitation versus Zen 2’s up to 64 cores, implying big gaps in high‑end and server scenarios.
  • Fabrication is seen as the main chokepoint: China lags in advanced nodes and lithography equipment, though SMIC is reported to be shipping 7 nm parts (e.g., Huawei phones, AI accelerators).

ISA, Architecture, and RISC‑V

  • Loongson now uses its own LoongArch ISA, derived from MIPS but modernized (no delay slots, no HI/LO, etc.).
  • SIMD extensions (LSX, LASX) resemble and extend earlier MIPS vector extensions.
  • The chip includes undocumented extensions to speed up binary-translated x86, ARM, and RISC‑V code.
  • Some commenters suggest Loongson should have gone with RISC‑V; others note RISC‑V wasn’t mature enough when LoongArch was defined, and its vector ecosystem is currently fragmented.

Software, OS, and Open Source

  • Linux support is advancing quickly; LoongArch changes for Linux, GCC, LLVM, and binutils are largely upstreamed.
  • Kernel drivers for the platform are mostly open source; one GPU driver is still out‑of‑tree, but available on GitHub, with only a closed‑source Mesa userspace component.
  • Windows support is considered unlikely, so Linux is seen as the natural desktop and server OS in China for this ISA.
  • Several point out that Chinese companies and academia are very active in open source (including RISC‑V projects) despite Western perceptions.

Geopolitics, Sanctions, and Strategy

  • Large subthread debates whether US chip sanctions are a strategic blunder or necessary containment.
  • One view: sanctions accelerate Chinese self‑reliance, destroy the “buy foreign, don’t build” mentality, and catalyze domestic ecosystems across chips, tools, and AI.
  • Counter‑view: even if China eventually catches up, delaying it 5–10 years is worthwhile for the US and allies.
  • Many frame this as part of a broader “tech war” and de‑globalization; opinions diverge sharply on long‑term outcomes.

Economics, Demographics, and Long‑Term Outlook

  • Some predict China will face a prolonged economic slump, property bust, and severe demographic issues that limit its ability to fund cutting‑edge fabs.
  • Others note that “China collapse” narratives have been repeated for decades and emphasize its large market, human capital, and existing R&D base.
  • There is broad agreement that design capability is strong in China; manufacturing equipment and advanced fabs remain the main constraint.