Better known among enthusiasts, the Chinese Loongson announced this week the 3D5000, its newest processor for data centers. Based on the proprietary LoongArch architecture, the solution uses two of the chiplets that made up the previous generation to reach 32 cores, and promises to be up to 4 times faster than competitors based on the ARM architecture — the same as cell phone chips and Apple Silicon —, having as the most important aspect to provide independence from foreign technologies to China’s servers.
The 3D5000 uses the LoongArch architecture, a Loongson proprietary ISA (Instruction Set Architecture, as the architectures are also called) released in 2020. LoongArch is based on RISC (Reduced Instruction Set Computer, or Reduced Instruction Set Computer), an architecture style that, as the name suggests, has a reduced number of limited instructions, faster to execute.
To give you an idea, the popular ARM, used in cell phones and Apple computers, is also based on RISC, while AMD and Intel processors use x86, an architecture based on CISC (Complex Instruction Set Computer), which uses instructions very more complex, but consequently more versatile — this is one of the aspects that most differentiates these chips and ends up determining their use.
With that out of the way, the 3D5000 is actually a combination of two units from its predecessor, the 3C5000, into a single package. There are two chiplets with 16 cores each, thus reaching a total of 32 LA464 cores, which, at the time they were announced, were aiming at the performance level of AMD’s Zen and Zen+ cores. The set works at a modest 2.0 GHz, carries 64 MB of L3 cache and features 8 memory channels for DDR4-3200 modules with error correction (ECC).
An interesting aspect is that the launch has automatic frequency and voltage adjustments, and thus, although it has a consumption established at 300 W, it usually operates at 150 W, according to Loongson data. There is also use of the little-discussed HyperTransport 3.0 bus (HT 3.0), used to allow configurations of two CPUs per server (2P) or even four CPUs per server (4P), complemented by the company’s other release, the 7A2000, a coprocessor dedicated to manage these settings.
The 7A2000 would be 400% faster than the previous generation, and would be responsible for ensuring the proper functioning of 2P and 4P servers, with up to 128 cores. Combined, the specifications would guarantee up to 1 TFLOP of computational power for data in FP64 format, common in tasks that require high precision, in addition to 2 TFLOPs in 2P machines and 4 TFLOPs in 4P servers, which would result in performance 4 times faster than the from rivals based on the ARM architecture, according to Loongson.
On the other hand, the published benchmark numbers are not very encouraging. The component would reach a score of 425 in the SPEC CPU 2006 benchmark, a common test solution for CPUs, but which has long been replaced by the more modern and demanding SPEC CPU 2017 — results of this version were not released. It is not hard to imagine that the recent AMD EPYC 9004 Genoa, with Zen 4 cores, and Intel Xeon Sapphire Rapids, with Golden Cove cores (those of the 12th generation for desktop) deliver better performance.
That said, the idea of the Loongson 3D5000 isn’t to outdo them, but to free Chinese companies from relying on foreign solutions like these. With constant sanctions applied by the US, and now by European and Asian countries, China has been forced to develop its own technologies to maintain its internal market. The novelty fits into this scenario, and has already been confirmed to be used by a wide range of local companies.
There are some more curious consequences of sanctions on the 3D5000: in addition to the proprietary architecture, the processor has custom mechanisms to circumvent security flaws such as Specter and Meltdown, which affected many generations of AMD and Intel CPUs, a TPM (Trusted Platform Module) ) and even a secret national encryption algorithm that would deliver a high efficiency of 5Gbps.
It is not common for manufacturers to announce prices for this type of solution, and in the case of Loongson, which will only offer the 3D5000 in China, the situation is no different. Nor should we see more in-depth independent tests that prove the performance of the chip, considering the limited access to the novelty.
Source: go Tom’s Hardware
