Zhaoxin's ZX-F/KX-7000/KH-40000 and beyond

[SarahKerrigan]

It is speculated that KX-7000 uses TSMC 6nm.

The official website of VIA Next shows x86 6nm testing, and VIA-related news also talks about 6nm tape-out and mass production next year.

VIA very well may have intended CNS to be an N6 IP, but Zhaoxin is on the Entity List and as far as I know, using TSMC is simply not an option - unless the US issued a waiver, which I've not seen a ton of indication of them doing.

 

[igor_kavinski]

I have a question. How much silicon on the die is dedicated to spying?

 

[Dubble]

VIA very well may have intended CNS to be an N6 IP, but Zhaoxin is on the Entity List and as far as I know, using TSMC is simply not an option - unless the US issued a waiver, which I've not seen a ton of indication of them doing.

According to some rumors, KX-7000 is manufactured using a variety of different process combinations, possibly SMIC N+2 / TSMC 6nm / HLMC 14nm / TSMC 16nm.
KH-50000 same that.

 

[igor_kavinski]

Makes sense. They've tons of engineers, a lot more than the West so guess they use them to ensure their domestic architecture can be fabbed on a variety of processes.

 

[kadasaki]

Hefei Zhaoxin is included in the entity list, but not Shanghai Zhaoxin. The two companies are unrelated.

The official website of VIA NEXT shows that 6nm has been tested for a year, and the use of 6nm or 7mm was mentioned in the video a few years ago.

 

[prosty_mirek]

Zhaoxin is not on Entity List.
And what if it will be? All that AMD-Thatic-Hygon JV is on it now, and AMD is still able to produce Hygon CPUs. Despite sanctions, they modified zen1 to 'zen2 level' (Hygon no.4 core) and Hygon filed some patents.
If Zhaoxin really uses 7/6nm, then only to bump clock, and maybe tests new VIA services. Century Avenue is only slighty modified Yongfeng, and KX-7000 is only 8 core, it can be produced on older/chinese nodes. Juiliazui revision "G" bumps now to 3,3 ghz on 16nm, and better power efficiency, than before.

Interesting notes form VIA 2022 annual report:

Develop x86 server platform products with high integration, high DRAM capacity, high I/O expansion, multi-core and high economic benefits, new architecture GPU core ultra-low powerconsumption thin notebook computer special processor.
2.5D, 3D chip packaging development
Develop 7nm, 6nm and other advanced process chip development and design technology
X86 processor platform: develop a new x86 architecture 16nm 8-core integrate AI accelerationcore processor with high integration, high performance, and high economic efficiency x86platform products.
Since the company lost significant influence on Shanghai Zhaoxin Semiconductor Co., Ltd. in March 2021, it stopped usingthe equity method for the company. From March 2021, the relationship between the company and Shanghai Zhaoxin SemiconductorCo., Ltd. was change from Affiliated enterprises to substantial related parties.
https://via-corporate.s3.us-west-2....te_governance/BOD_2022/2022+Annual+Report.pdf

CHA/nCore is alive?

 

[Dubble]

Zhaoxin is not on Entity List.
And what if it will be? All that AMD-Thatic-Hygon JV is on it now, and AMD is still able to produce Hygon CPUs. Despite sanctions, they modified zen1 to 'zen2 level' (Hygon no.4 core) and Hygon filed some patents.
If Zhaoxin really uses 7/6nm, then only to bump clock, and maybe tests new VIA services. Century Avenue is only slighty modified Yongfeng, and KX-7000 is only 8 core, it can be produced on older/chinese nodes. Juiliazui revision "G" bumps now to 3,3 ghz on 16nm, and better power efficiency, than before.

Interesting notes form VIA 2022 annual report:

Develop x86 server platform products with high integration, high DRAM capacity, high I/O expansion, multi-core and high economic benefits, new architecture GPU core ultra-low powerconsumption thin notebook computer special processor.
2.5D, 3D chip packaging development
Develop 7nm, 6nm and other advanced process chip development and design technology
X86 processor platform: develop a new x86 architecture 16nm 8-core integrate AI accelerationcore processor with high integration, high performance, and high economic efficiency x86platform products.
Since the company lost significant influence on Shanghai Zhaoxin Semiconductor Co., Ltd. in March 2021, it stopped usingthe equity method for the company. From March 2021, the relationship between the company and Shanghai Zhaoxin SemiconductorCo., Ltd. was change from Affiliated enterprises to substantial related parties.
https://via-corporate.s3.us-west-2....te_governance/BOD_2022/2022+Annual+Report.pdf

CHA/Ncore may no longer be available, cns has more possibilities to be attacked by Spectre than yongfeng

 

[prosty_mirek]

I have found something interesting.
From "Shanghai Lianhe Investment Co., Ltd. 2023 Public Issuance of Technology Innovation Corporate Bonds (Second Series) Prospectus for Professional Institutional Investors" (PDF file, translated by ChatGPT):

Company I continues to innovate towards international advanced levels and technological frontiers, advancing the research and development of the new product CHX005. The CHX005, utilizing a 6nm process, will be coordinated with Chiplet to simultaneously support various product forms, including low-power notebooks, mainstream mid-range desktops, workstations, high-performance general servers, etc. Its single-core and overall performance are comparable to mainstream international products of the same period.

According to the company's development plan, Company I will further enhance core performance, optimize power management architecture, explore new possibilities in multi-core interconnect architecture, and research multi-layer high-performance interconnect architectures such as improved interconnects between cores, nodes, and sockets. The company will maintain a leading position in high-speed IO domestically, closely following the latest international specifications, fully supporting state-of-the-art computer interface standards. The company will also upgrade designs for DDR5 interface, PCIe4 interface, USB4 interface, and VT-D 2.5, achieving advanced IO productization.

In terms of next-generation research, Company I will continue to enhance core performance, aiming to increase IPC by 1.3 times. The company will explore higher frequencies, superior interconnect architectures, and conduct research on advanced process design technologies, breakthrough studies in low-power design, physical implementation, testability design, and other critical technologies. Simultaneously, targeting the vast and diverse market, the company will develop competitive high-performance and cost-effective products, further enhancing the core competitiveness of Company I's chip series products and providing sufficient technical support for gradually capturing more market share.

KX-6000 is HX002.
KH-40000 is HX003.
KX-6000g is HX004.
So KX-7000 should be HX005, and more indicates that it is actually 6nm process (and maybe 7nm for IO/GPU chiplet). Yet it is laughable if they squeeze only 3.7 ghz from 6nm, although promised "4.x ghx" in PPT. Or SMIC N+2 could be called 6nm?

Statement about "next-generation (...) increase IPC by 1.3 times" is also interesting. It can't be about Century Avenue, because that is already 2x Luijiazui and <10% up from Yongfeng. So it should be about KH-50000 and that puts it between zen2-3.

CHA/Ncore may no longer be available, cns has more possibilities to be attacked by Spectre than yongfeng

That should't be the case, why Zhaoxin not choose CNS. Vendors were aware of Spectre since 2017. Centaur unveils CNS/nCore in 2019. VIA offered it until at least 2022. Luijiazui is also vulnerable and Zhaoxin relaunch it few years later. CNS was hot at higher frequencies (flawed implementation?), and nCore AI benefits from AVX512. New cores do not have AVX512 yet.

 

[Dubble]

I have found something interesting.
From "Shanghai Lianhe Investment Co., Ltd. 2023 Public Issuance of Technology Innovation Corporate Bonds (Second Series) Prospectus for Professional Institutional Investors" (PDF file, translated by ChatGPT):


KX-6000 is HX002.
KH-40000 is HX003.
KX-6000g is HX004.
So KX-7000 should be HX005, and more indicates that it is actually 6nm process (and maybe 7nm for IO/GPU chiplet). Yet it is laughable if they squeeze only 3.7 ghz from 6nm, although promised "4.x ghx" in PPT. Or SMIC N+2 could be called 6nm?

Statement about "next-generation (...) increase IPC by 1.3 times" is also interesting. It can't be about Century Avenue, because that is already 2x Luijiazui and <10% up from Yongfeng. So it should be about KH-50000 and that puts it between zen2-3.


That should't be the case, why Zhaoxin not choose CNS. Vendors were aware of Spectre since 2017. Centaur unveils CNS/nCore in 2019. VIA offered it until at least 2022. Luijiazui is also vulnerable and Zhaoxin relaunch it few years later. CNS was hot at higher frequencies (flawed implementation?), and nCore AI benefits from AVX512. New cores do not have AVX512 yet.

Well, maybe it is confirmed that Zhaoxin has a better architecture, the zx-f/cns consumes 140W at 2.5GHz avx, and at 2.2GHz, only 65W.

 

[prosty_mirek]

Well, maybe it is confirmed that Zhaoxin has a better architecture, the zx-f/cns consumes 140W at 2.5GHz avx, and at 2.2GHz, only 65W.

CNS was rushed. Centaur states before that because of schedule, they don't implement SMT and uop cache. Indeed it had frequency, power and temperature issues, but not necessarily because of architecture, but from in silicon implementation. In 2019 it was only engineering sample. On that time VIA group don't even had patent mentioning AVX-512. Via Alliance/Zhaoxin team filed for it in 2020/2021.

 

[NostaSeronx]

Zhaoxin is banned from using TSMC because they are dual-use. Which means they fall under the shared US/Taiwan export restrictions:
"additional license requirements for microprocessors having a processing speed of 5 GFLOPS or more and an arithmetic logic unit with an access width of 32 bit or more, including
those incorporating “information security” functionality, and associated “software” and “technology” for the “production” or “development” of such microprocessors."
TSMC only has this license for Nanjing/Fab 16.

"PLA procurement tenders require that systems support Zhaoxin chips, like the Strategic Support Force’s call for bids for a data management system on
behalf of PLA Unit 61081 that requires that the system support domestic chips like Feiteng, Longsun, Shenwei, or Zhaoxin."

2019: ordered all state offices to remove foreign hardware and software within three years
this was extended in 2022 to a by ~May 2024 date. Assume this will get its completion extended but will start with KX-7000/KH-50000 for x86 government desktops/servers.

If Zhaoxin wants its PLA/government contract they need to fabricate at SMSC/SMIC. Of which, VIA Next has SMIC labeled in foundries: https://www.vianextech.com/competencies/

NDA over node used could be like a N+2 Enhanced/Plus/slight shrink or whatever. Since, I don't see why TSMC requires a NDA. Unless they know they are going to get in trouble for giving N6 to Nanjing.

The path of least resistance is the domestic variant with Zhaoxin uses SMIC and the global variant with VIA uses TSMC.

From VIA's 2022 annual report:
VIA TECHNOLOGIES, INC. owns 100% of VIA NEXT TECHNOLOGIES, INC.
"In addition to accelerating its momentum in China, VIA NEXT will continue to expand and develop new customer groups in other markets."

Taipei => Handles VIA's return to self-made chips instead of using Intel/Qualcomm/MediaTek/NXP. (Other markets)
Shanghai => Continues to support Zhaoxin. (China)

 

[Kosusko]

CentaurHauls Family 7 Model 11 Stepping 1
Zhaoxin Kaixian KX-7000’s GeekBench scores revealed, performance doubled compared to previous generation
source: https://article.pchome.net/article/content-2173598.html

 

[AnitaPeterson]

CentaurHauls Family 7 Model 11 Stepping 1
Zhaoxin Kaixian KX-7000’s GeekBench scores revealed, performance doubled compared to previous generation
source: https://article.pchome.net/article/content-2173598.html

I see you're also running one of these chips, in a dedicated motherboard combo. Checked it out, it's quite wild: Laptop memory, regular ATX power, lots of USB, 4 SATA ports... And you're running Windows 10 on it? How does it perform? Are there any glitches or errors in programs or games?

 

[igor_kavinski]

How does it perform?

I think that's his system running Doom (2016):

 

[eprenrem]

IIRC, Cyrix/Via has x86 IP, however that is the old stuff. I don't believe they have any right to use anything new Intel/AMD has developed after that. Thats all Intel/AMD's own IP.

 

[SarahKerrigan]

IIRC, Cyrix/Via has x86 IP, however that is the old stuff. I don't believe they have any right to use anything new Intel/AMD has developed after that. Thats all Intel/AMD's own IP.

VIA had a capture agreement with Intel extending well into the 2010s. Hence, AVX512 support in CNS.

 

[Dubble]

KX-7000/8 detail by china tester
8-Core CCD die size ≈ 78.5mm2
IOD die size ≈ 152mm2.

not boost
all cores 3.2GHz
single core 3.6GHz

boost
all cores boost 3.6GHz
single core boost 3.8GHz
after some day maybe reach 4GHz

 

[raising19]

KX-7000/8 detail by china tester
8-Core CCD die size ≈ 78.5mm2
IOD die size ≈ 152mm2.

not boost
all cores 3.2GHz
single core 3.6GHz

boost
all cores boost 3.6GHz
single core boost 3.8GHz
after some day maybe reach 4GHz

not boost
cpuz
all cores 3.1GHz
single core 3.3GHz
7zip
all cores 3.0GHz

boost
all cores 3.6GHz
single core 3.6GHz