How big is the gap between domestic x86 CPU and Intel? Less than half the performance of mainstream quad-core products

In June this year, Wei Shaojun, chief engineer of the National Nuclear Corporation, said in an interview that "the successful independent research and development and mass production of the domestic Zhaoxin x86 general-purpose processor has enabled the performance of domestic desktop processors to achieve a leap-forward improvement, from less than 7% of the international overall level at the beginning of the 12th Five-Year Plan to the current 50%, and the comprehensive performance experience has reached 80%." In fact, how does the performance of domestic X86 chips compare with Intel? What progress has been made in the process of technology introduction, absorption and innovation?

A few months ago, Shanghai Zhaoxin's latest product ZXD was unveiled at the 2017 Beijing International Internet Technology Expo and World Cyber ​​Security Conference. Although ZXD is still far behind Intel and AMD's CPUs, it is indeed a great improvement compared to Zhaoxin's first CPU ZXA. Since Shanghai Zhaoxin was founded in 2013, from the first product ZXA to the latest product ZXD, Shanghai Zhaoxin has spent several years to improve the bare CPU performance of its products by more than 80% based on VIA's technology.

How much performance improvement does ZXC have compared to ZXA?

ZX-A is Zhaoxin's first CPU, with technology derived from VIA, 2 cores, 1.6G main frequency, and TSMC 40nm manufacturing process. In terms of its performance, according to the SPEC2006 test of relevant units, the 1.6G main frequency ZXC with Ubuntu 14.04 operating system and ICC compiler has the following results:

SPECint2006:11.3

SPECint_rate2006: 20.5

SPECfp2006: 12.4

SPECfp_rate2006:18.7

The single-thread scores of the STREAM test are:

Copy 5502.8MB/s;

Scale 5042.9MB/s;

Add 5321.6MB/s

Triad 5252.7MB/s

The multi-threaded scores of the STREAM test are:

Copy 5191.1MB/s;

Scale 5104MB/s;

Add 5383.8MB/s

Triad 5461.9MB/s

Here is an introduction, SPEC2006 is a relatively authoritative test software in the industry, it tests the CPU through 26 programs, and then calculates the results, the higher the score, the better the performance. However, the SPEC2006 test is not flawless - you can cheat by not specifying the compiler and other conditions, whether or not to open auto parallelization will also have a difference in the score, and there will be a certain difference between the base score and the peak score, etc.

Even so, compared with some black box tests, SPEC is still a relatively fair test that can objectively reflect the actual performance of the CPU. After all, there is no absolutely objective test, just like the college entrance examination. Even if there are problems of one kind or another, it is still the most operational and relatively fair and reasonable evaluation method (except Beijing and Shanghai).

ZX-C is an introduction and digestion of VIA technology. Compared with ZXA's 2 CPU cores, ZXC has 4 CPU cores, and the CPU cores interact with each other through L2 Cache. Compared with ZXA, ZXC's multi-core performance has been improved. In view of the resource redundancy and imbalance of the core introduced from VIA, it has been redesigned and TSMC's 28nm manufacturing process has been adopted to reduce power consumption, improve stability, reduce mass production costs, and increase the CPU's main frequency from ZXA's 1.6G to ZXC's 2G.

In addition, Zhaoxin also redesigned the FSB interface, and the FSB frequency was increased from 1066MHz to 1333MHz. The memory access unit was redesigned, and a new L1/L2 memory access control unit was designed to optimize memory access performance.

The most direct manifestation of these modifications is in performance. According to the SPEC2006 test conducted by relevant units, the ZXC with a main frequency of 2.16G (Turbo) using the Ubuntu 14.04 operating system and the ICC compiler has the following results:

Specint 2006:17

Specint rate2006：50

Specfp 2006:18.2

Specfp rate2006：36.3

Compared with ZXA, ZXC has improved its single-threaded fixed-point performance by about 54% and its multi-threaded performance by about 147%. In terms of floating-point performance, ZXC has improved its single-threaded performance by about 50% and its multi-threaded performance by about 50%.

Although in the case of Ubuntu 14.04 operating system and ICC compiler, the results of Specint 2006:17; Specfp 2006:18.2 are far behind those of Intel, ZXC has made great progress compared to ZXA.

ZXD performance is further improved and makes up for the biggest shortcoming of ZXC

As mentioned before, ZXC has improved single-thread fixed-point performance by about 54% and multi-thread performance by about 147% compared to ZXA; in terms of floating-point performance, ZXC has improved single-thread performance by about 50% and multi-thread performance by about 50% compared to ZXA.

But ZXC also has its own shortcomings, that is the front-end bus - for applications such as images, audio, video, games, FPGA synthesis and scientific applications, which usually operate on a small part of the data of a large working set, the front-end bus becomes a major performance bottleneck.

ZXC continues the design of VIA's original front-side bus (VIA V4 bus) - the front-side bus (FSB) is the computer communication interface of early Intel chips, similar to AMD's EV6, which connects the CPU and the north bridge chip. The memory controller is usually integrated in the north bridge - ZXC did not integrate the memory controller into the processor, but in the bridge chip. As a result, the results of the STREAM test are not very good.

According to the results of the STREAM 5.10 test, the single-threaded results are:

Copy：7685.0 MB/s

Scale: 7446.4 MB/s

Add: 8129.5 MB/s

Triad: 7981.7 MB/s

The 4-thread test results are:

Copy: 7426.3 MB/s

Scale: 7504.0 MB/s

Add: 7656.8MB/s

Triad: 7787.7 MB/s

From this, we can see that the STREAM test results of ZXC are relatively average, and the improvement is relatively limited compared to ZXA.

The most obvious improvement of ZXD is the replacement of a better DDR4 memory controller. It is the first domestic desktop CPU with an integrated DDR4 memory controller. Moreover, ZXD has integrated the memory controller into the chip. Compared with ZXC's DDR3 memory controller, ZXD uses a DDR4 memory controller and integrates the memory controller into the chip, rather than integrating the memory controller into the bridge chip like ZXC. This change has improved the overall performance of the CPU to a certain extent.

In addition, ZXD is likely to have replaced the FSB bus with an interface similar to Intel's DMI. The combination of these two factors eliminates the shortcomings of ZXC. The most typical proof is the significant improvement in STREAM test scores. According to the tests of relevant units, the single-thread scores of STREAM 5.10 tests are:

Copy: 10942.4 MB/s

Scale: 10371.4 MB/s

Add: 10603.1 MB/s

Triad: 9850.1MB/s

The multi-threaded results are:

Copy: 12666.1 MB/s

Scale: 13060.3 MB/s

Add: 11270.2MB/s

Triad: 11302.6MB/s

From the test results, we can see that ZXD's SRTEAM test results have been significantly improved compared to ZXC, making up for the shortcomings in the past. And by updating the DDR4 memory controller, it can also effectively deal with the problem of DDR3 memory sticks being discontinued by major memory manufacturers in business.

In addition to the improvements on the memory controller, ZXD adds an X86 instruction buffer, provides a more accurate loop buffer function, and significantly reduces the number of pipeline stages through the optimization of the front and back ends of the overall pipeline, effectively reducing the performance loss of branch prediction failures... After a series of improvements, it is directly reflected in the performance improvement of ZXD compared to ZXC. According to the data of relevant units, when using the Ubuntu 14.04 operating system, 32GB DDR4 memory, Intel 530 series 120GB SSD hard disk, ICC compiler, and ZXD turbo frequency to 2.2G, the SPEC2006 test results are:

Specint 2006: 20.4

Specint rate2006：63.3

Specfp 2006:23

Specfp rate2006：47.6

Although ZXD's performance is still far behind Intel and AMD, in terms of bare CPU performance, ZXD is about 30%-40% of Intel's mainstream quad-core CPU. And it must be pointed out that Zhaoxin's modification of the CPU architecture is relatively limited, and the current modifications are not a complete overhaul, for example, it has not achieved the same improvement and leap as AMD from "pile driver" to Zen.

As a winner of Toutiao's Qingyun Plan and Baijiahao's Bai+ Plan, the 2019 Baidu Digital Author of the Year, the Baijiahao's Most Popular Author in the Technology Field, the 2019 Sogou Technology and Culture Author, and the 2021 Baijiahao Quarterly Influential Creator, he has won many awards, including the 2013 Sohu Best Industry Media Person, the 2015 China New Media Entrepreneurship Competition Beijing Third Place, the 2015 Guangmang Experience Award, the 2015 China New Media Entrepreneurship Competition Finals Third Place, and the 2018 Baidu Dynamic Annual Powerful Celebrity.