The Geekbench 5 synthetic test database introduced data on the KaiSheng KH-40000 processors' performance from China's Zhaoxin company. A formal set of chips has not yet been submitted. It will include models of up to 16 kernels. The first tests show that, against the background of the predecessors, the KaiSheng KH-40000 has grown, but it does not reach modern AMD and Intel desktop processors.
Zhaoxin, owned by Via Technologies and the Shanghai government, has continued to use Via's micro-architectures in its products since the middle of the last decade.
The Geekbench 5 test included a 16-nuclear model of KaiSheng KH-40000/16 and a 12-nuclear model of KaiSheng KH-40000/12. Both chips operate at a frequency of 2.20 GHz. It is known that the older model has 32 MB-Kash-L3, the younger one 24 MB-Kesh-L3. The older model supports SMT technology, so the processor has 32 virtual streams. Based on data collected by the Geekbench 5 test, the Zhaoxin KaiSheng model KH-40000/16 and the KaiSheng KH-40000/12 are very similar to the unreleased Centaur CHA processor reported earlier. However, there is a difference between chips. It is half the number of kernels collected by the Centaur CNS, it has no support for SMT, as well as the use of the more mature process process of TSPM production, the new THMC N40000.
Although the new KaiSheng KH-40000 CentaurHaul processors operate at a frequency of 2.20 GHz, they are much faster than the previous generation ' s chips on the LuJiazui architecture, which operate at a higher frequency of 2.70 GHz. This is true for operations involving numerical values where new chips are 22% faster, as well as those with floating point numbers where their productivity is 75% higher.
Against the background of the predecessors, represented in a comparative test by the eight-nuclear model Zhaoxin KX-U6780A, the KaiSheng KH-40000 processors are much better; however, compared to any modern AMD and Intel processors, they are playing clean. For example, their single-track performance is even lower than that of the AMD chips on Bulldozer/Piledriver architectures in 2012.
The difference between the 12th and 16th nuclear models of KaiSheng KH-40000 is also very significant. In the theory of a chip with a support of 32 streams, the 32 core actually can handle 2.66 times the flow of a processor that has 12 cores with 12 streams. However, as shown above, KaiSheng KH-40000/16 is even faster: 2.69 times the Integer task and 2.84 times the Float task. The senior processor has only four more physical cores than the youngest, but its productivity is almost three times higher.
It should be noted that Windows 10 and Windows 11 operating systems do not always work optimally with the task planner with the participation of an unknown multi-nuclear processor. Therefore, the results of the 12-nuclear model KaiSheng KH-40000/12, which was tested in Windows 10 Pro, may not reflect the real potential of the processor against the background of the older model, which was tested in a more appropriate environment. Even so, the multi-nuclear productivity of the 12-nuclear model on the CentaurHoals architecture was expected to be higher than that of the chip on the previous LuJazui architecture. The latter has not only a smaller number of kernels, but also uses the older process and has a smaller volume of Cache memory L3. In operations Integer and Float, the difference in favour of the newer model was 40 per cent and 78 per cent respectively. However, as mentioned above, in general, the speed of the KaiSheng KH-40000 processors and their predecessor Centau CHA are very high on modern standards.
For example, the performance of the 12-nuclear KaiSheng KH-40000/12 in Windows is comparable to the 8-nuclear AMD FX-8350.