For years, the most repeated idea in the sector was simple: if the United States cut off China’s supply of advanced chips and critical machinery, the country would be condemned to “play in a different league,” with slow, uncompetitive domestic processors. However, the scenario has become more challenging for the West: Beijing has accelerated investments, pushed its companies to build alternatives, and occasionally, signs of real progress emerge. One of these now comes from the most everyday realm —the PC— with the first public analyses of the Loongson 3B6000, a 12-core processor that aims to demonstrate that “Made in China” computing is not limited to low-cost products or consumer electronics.
The context matters. U.S. restrictions on advanced semiconductors and related technologies, active since October 2022 and tightened in subsequent years, have pressured China to reduce dependence on external sources for high-performance computing. This tension has also spilled into public debate with media milestones like DeepSeek-R1, which in 2025 fueled the narrative of a Chinese leap in AI models and shook the market: the announcement triggered a wave of sales in AI-related stocks and hit NVIDIA hard in a particularly tough session, as reported by agencies and financial media.
But one thing is the race of models, and another, slower and more costly, is the race of general-purpose hardware. That’s where Loongson enters—one of the historic names in China’s CPU industry, which has been evolving for years from previous designs and a strategy of technological independence based on its own architecture. The 3B6000 is not an x86 chip like Intel’s or AMD’s: it relies on LoongArch, a proprietary instruction set that requires traversing a steeper path in compatibility, compilers, libraries, and optimizations. In other words: having “more cores” isn’t enough; there needs to be an ecosystem that truly leverages those cores.
What benchmarks say: better than a “Raspberry Pi on steroids,” far from a modern PC
The tests sparking the debate come from extensive benchmarking on Linux. The Loongson 3B6000 test setup included 64 GB of DDR4 at 3,200 MT/s and a Samsung 980 Pro 2 TB NVMe SSD to eliminate bottlenecks unrelated to the CPU. The comparison included current processors from AMD (Zen 5, like the Ryzen 5 9600X) and Intel (Arrow Lake, like the Core Ultra 200K), along with a Raspberry Pi 500+.
The overall result is quite clear: the Loongson 3B6000 outperforms the Raspberry Pi 500+, but is well behind any modern AMD or Intel CPU in most scenarios. In the geometric mean of the battery — a way to summarize dozens of tests without a single one dominating — the 3B6000 ranks towards the bottom with 248 points, compared to 102 points for the Raspberry Pi 500+ (roughly 2.5 times less), but still far from the 775 points of the Ryzen 5 9600X or the 824 points of the Core Ultra 5 245K. At the top of the list, a Ryzen 9 9950X3D achieved 1,352 points in the same summary.
When examining details, the pattern repeats with interesting nuances. In compilation, a very real developer workload, the 3B6000 takes 65.85 seconds in Timed Erlang/OTP, ahead of the 163.23 seconds of the Raspberry Pi 500+, but far behind the 24.37 seconds of the Ryzen 5 9600X. In Timed PHP, the difference persists: 125 seconds for the Loongson, versus 440 seconds in the Raspberry Pi and 54 seconds in the 9600X.
And then comes the part that explains why these chips still do not “compete” in the Western sense of the term. In workloads with highly optimized code for x86_64 — especially with vector extensions like AVX-512 in AMD and finely tuned assembly routines — the Loongson struggles. For example, in x265 video encoding, performance is described as up to 10 times lower than the Ryzen 5 9600X in published tests. It’s not just about raw power: it’s about software maturity, libraries, and years of micro-optimizations on a specific architecture.
The “surprises” that nuance the story
If the 3B6000 were simply “slow,” the discussion would end quickly. What’s striking is that some tests do show signs of competitive potential.
In C-Ray 2.0, a classic ray tracer, the Loongson 3B6000 matched the Ryzen 5 9600X. This is an uncomfortable comparison: the Chinese chip doubles cores (12 vs. 6), yet the tie is relevant because it shows that when code is relatively portable and scales well across threads, the processor can hold its own and defend,”
There are also notable results in Quicksilver, where the 3B6000 performed at the level of a Core Ultra 9 285K in the cited test, and in BYTE UnixBench 6.0.0, where it was close to the Ryzen 5 9600X. These are “islands” within a sea of unfavorable benchmarks but help convey Loongson’s message: the chip isn’t a toy; it’s a serious attempt to build a sovereign PC platform, even if today it’s unbalanced.
Why does this happen: frequency, DDR4 memory, and the “cost” of a proprietary ISA
Part of the explanation is physical: according to technical analyses shared after testing, the Loongson 3B6000 operates around 2.5 GHz, a figure well below typical (and peak) clocks in modern Intel and AMD desktop CPUs. Additionally, the tested platform relies on DDR4-3,200, while current PC markets push towards DDR5 with higher bandwidths.
But the decisive factor is software. Intel and AMD have decades of optimizations for x86 in compilers, runtimes, and libraries. Many “industrial” workloads (from multimedia to encryption) include specific routines optimized for x86_64. LoongArch, by definition, starts at a disadvantage: it needs time, community, tools, and market volume to develop these optimizations.
The system setup and accompanying hardware also show that this is not a “consumer” platform in the usual sense. The tested motherboard has two DDR4 slots and specific recommendations for registered ECC memory; overall, the ecosystem outside China remains limited. There’s also a significant gap for energy analysts: Linux lacked support to measure CPU consumption with the accuracy of modern AMD and Intel platforms, making it difficult to compare efficiency rigorously.
Can it compete with the West? It depends on how “compete” is defined
If “competing” means matching performance, efficiency per watt, and user compatibility — gaming, content creation, workstations, commercial software — the answer today is no. In most tests, the 3B6000 is below even current mid-range CPUs, despite having 12 cores.
But if “compete” means something else —building a viable alternative for a massive domestic market, reducing strategic dependence, and sustaining an indigenous industrial chain— then the 3B6000 is indeed a relevant piece. China, in this scenario, doesn’t need to win the “YouTube benchmark” to consider the effort successful: it suffices that the platform is sufficiently useful, scalable, and controllable.
And that’s the key: the Loongson 3B6000 may not be the chip to dethrone Intel and AMD in desktops, but it’s a reminder that the technological gap can be narrowed through sustained investment… even if the final performance depends both on silicon and the software that runs on it.
Frequently Asked Questions
What is LoongArch, and why does it complicate comparison with Intel and AMD?
LoongArch is a proprietary instruction set architecture. Since it’s not x86_64, it doesn’t inherit decades of optimizations and direct compatibility with software compiled for Western PCs, impacting real-world performance and application availability.
What kind of workloads make sense for a CPU like the Loongson 3B6000?
Primarily for environments prioritizing technological sovereignty, controlled deployments, software specifically compiled for the platform, and workloads where the ecosystem —not just raw power— is a strategic requirement.
Why does performance drop so much in x265 or certain libraries?
Because many popular tools include specific optimizations for x86 (and advanced vector extensions). On less common architectures, those paths don’t exist or are less mature, making performance rely on more generic implementations.
What would need to happen for Loongson to approach the performance of a modern PC?
Beyond architectural and frequency improvements, the LoongArch ecosystem needs to grow: better compilers, optimized libraries, more polished kernels and distributions, and, most importantly, sufficient adoption to justify years of optimization work.