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

Page 7 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,542
14,496
136
You're desperately trying to defend this processor's performance, comparing to your FX-8300 under-clocked to 2.7 GHz saying how great it would be for casual office use and light gaming.

The FX series was one of AMD's biggest blunders, and early chips like Bulldozer and Piledriver had worse IPC than the Phenom II predecessor released in 2008. This 2020 CPU has IPC equivalent to that of CPU designs released 12 years ago, is running at clock speeds from 12 years ago and is consuming power like CPUs from ~8 years ago.

I had an FX processor and it wasn't great at games at 4.1 GHz - at 2.7 GHz it'd be downright miserable. The only thing the Zhaoxin has going for it is lots of cores.
@Kosusko , This is what I am talking about. I know you don't have recent hardware, that not your fault, as you most likely don't have the budget for it.

The problem is, that you keep trying to show all these benchmarks of slow hardware. Its not exciting. Just read the foums and be happy for the day you can but something new, fast and exciting, but posting benchmarks of old or slow hardware is just annoying.
 
Apr 30, 2020
68
170
76
The performance would be tolerable for an office machine, but I suspect even 1080p60 YouTube would struggle without GPU acceleration. Games from the 2008-2010 Era would probably run fine, but games from 2016+ demand way too much individual core performance.

The thing that really makes it unacceptable is the power consumption. If this thing had a 5w TDP - we'd have something. But at 70w it's just bizarre.
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
I think the Zhaoxin benchmarks are interesting from a purely academic standpoint. Of course, no one here is actually going to buy one, nor does anyone have a reason to, but it doesn't hurt to keep tabs on them.
 
  • Like
Reactions: lightmanek

ThatBuzzkiller

Golden Member
Nov 14, 2014
1,120
260
136
I have no idea how hard they work their engineers. I can only evaluate them based on their products and their release schedule.

The working conditions are far more worse at Zhaoxin than wherever the CPU design teams are based at Intel. Labour laws are nonexistent where Zhaoxin is located and with such a government deemed critical project, I doubt the lights go out for a lengthy period if ever at all ...

Their release schedule seems like they're slowing down. I thought they were supposed to be off the old VIA cores by now? We've been hearing about the KX-7000 for a long time.

They opened in 2013 and nearly none of the staff had previously architected VIA's designs so it was going to be an early upbringing in their state-backed corporate life. No idea about the tooling situation at the time either and then there's other stuff like the quality of the documentation as well which can vastly effect the release schedule too. It's a very hard to make the argument that they were initially planning to develop the KX-7000 early on shortly after their opening ...

I can only see their release schedule accelerating given the deteriorating political situation which will further tempt their government sponsor to pour in more resources into development ...
 

scannall

Golden Member
Jan 1, 2012
1,946
1,638
136
@Kosusko , This is what I am talking about. I know you don't have recent hardware, that not your fault, as you most likely don't have the budget for it.

The problem is, that you keep trying to show all these benchmarks of slow hardware. Its not exciting. Just read the foums and be happy for the day you can but something new, fast and exciting, but posting benchmarks of old or slow hardware is just annoying.
Well, now you can get great performance cheap with the 3300X. ;-) There is hope in the future.
 

ThatBuzzkiller

Golden Member
Nov 14, 2014
1,120
260
136
Which one ?

Working conditions ? I could imagine fewer willing highly capable participants joining their projects.

Documentation ? This is practically par for course with many older projects since they are not as well documented as projects are today so I can see some time being dedicated to reverse engineering just to fix past design flaws because they can't rely on getting information from previous architects.
 

Kosusko

Member
Nov 10, 2019
161
120
116
CentaurHauls Family 7 Model 11 Stepping 1 vs CentaurHauls Family 6 Model 71 Stepping 1
Zhaoxin KX-U6780A 2.7GHz vs Maybe Zhaoxin KH-40000 Series 2.0GHz "ES"
8 small cores vs 8 BIG cores

Single-Core Score: 362 (77,19%) vs 469 (+29,56%)
Multi-Core Score: 2387 (73,13%) vs 3264 (+36,74%)


_________________

New desktops and notebooks also ultra-thin notebooks on the way:


The competition is coming to the world of duopoly....
 
Last edited:
  • Like
Reactions: lightmanek

NostaSeronx

Diamond Member
Sep 18, 2011
3,686
1,221
136
Zhaoxin KH-40000
Too bad the KX-40000 isn't coming out this year eh?
The KH-40000 is the big 16-core die and it uses the same core as ZX-E/KX-6000 series. This one is paired with the discrete GPU also built by Zhaoxin.
=> 4x 16-core dies in two packages ==> 64-core 2p bga motherboard.

ZX-F is used in KX-7000 and KH-50000.

Only way for ZX-F would be good is if its TDP was really low at 2 GHz or it will eventually clock at >4 GHz.
 
Last edited:

Kosusko

Member
Nov 10, 2019
161
120
116
The KH-40000 is the big 16-core die and it uses the same core as ZX-E/KX-6000 series. This one is paired with the discrete GPU also built by Zhaoxin.

ZX-F is used in KX-7000 and KH-50000.


2020 - 16nm FFC TSMC KaiXian KX-6000 Series (ZX-E) for office computers / multimedia
2020 - 16nm FFC TSMC KaiSheng KH-30000 Series (ZX-E) up to 8 cores on one die and dual socket solution up to 16 cores

2021 - 16nm FFC TSMC KaiSheng KH-40000 Series (ZX-F) up to 8 cores on one die, two die 16 cores and dual socket solution up to 32 cores for server computers

2022 - 7nm KaiXian KX-7000 Series up to 16 cores for gaming computers
2022 - 7nm KaiSheng KH-50000 (ZX-G) up to 16 cores on one die, two die 32 cores and dual socket solution up to 64 cores for server computers


ZX_G roadmap.png



ZX_F_ZX_G roadmap.png
 
Last edited:

ThatBuzzkiller

Golden Member
Nov 14, 2014
1,120
260
136
They are on the right path in an industry where progress is not an absolute given anymore ...

Sooner or later all of them will find themselves in disillusionment but some like Intel are experiencing this faster while others like Zhaoxin stand to gain the most with the coming of a post-silicon future where improvements to transistor technology and performance will be few and far in between ...

Deadlines aren't the ultimate endgame, it's about raising performance targets ...
 

dark zero

Platinum Member
Jun 2, 2015
2,655
138
106
@Kosusko , This is what I am talking about. I know you don't have recent hardware, that not your fault, as you most likely don't have the budget for it.

The problem is, that you keep trying to show all these benchmarks of slow hardware. Its not exciting. Just read the foums and be happy for the day you can but something new, fast and exciting, but posting benchmarks of old or slow hardware is just annoying.
Maybe is to show up how the proccesor evolved in the latests years. Before that VIA/ZhaoXin literally was nothing. Even ARM A53 cores have more potential than it. Now at least tries to reach the numbers of the infamous AMD FX, but is matter of time to get better and better.
The question is on how many generations they will be decent.
 

diediealldie

Member
May 9, 2020
77
68
61
  • Like
Reactions: lightmanek

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,830
136
Wow...somehow Thuban is looking great again. Even with extremely good AES results which uses newer instruction sets, chinese CPU cannot stand against Thuban, even if it's scaled to(2.7/4.0) same clock.

The problem is that Zhaoxin is still stuck to their VIA roots, using what is essentially an old core in a new package. And that core wasn't necessarily that great to begin with.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,686
1,221
136
The problem is that Zhaoxin is still stuck to their VIA roots, using what is essentially an old core in a new package. And that core wasn't necessarily that great to begin with.
Actually, Zhaoxin left VIA with ZX-D. The government was suspicious about ZX-C and older cores. So, the cores for ZX-D and ZX-E are homegrown with their own optimizations. However, they haven't improved the execution core component that much. Lack of ALUs/FPUs/AGUs is a major issue, with the front-end and back-end being pretty much on-par with CNS. Which imo Zhaoxin never improving; the amount of execution units or the load-store sub-system, being slightly weird.
 
Last edited:

Kosusko

Member
Nov 10, 2019
161
120
116
The problem is that Zhaoxin is still stuck to their VIA roots, using what is essentially an old core in a new package. And that core wasn't necessarily that great to begin with.

Actually, Zhaoxin left VIA with ZX-D. The government was suspicious about ZX-C and older cores. So, the cores for ZX-D and ZX-E are homegrown with their own optimizations. However, they haven't improved the execution core component that much. Lack of ALUs/FPUs/AGUs is a major issue, with the front-end and back-end being pretty much on-par with CNS. Which imo Zhaoxin never improving; the amount of execution units or the load-store sub-system, being slightly weird.

VIA CPU = Centaur Technology design and Zhaoxin CPU = Centaur Technology design.
You spin in a circle.

 

NostaSeronx

Diamond Member
Sep 18, 2011
3,686
1,221
136
VIA CPU = Centaur Technology design and Zhaoxin CPU = Centaur Technology design.
VIA x86 CPU = Centaur
Zhaoxin x86 CPU = Zhaoxin(ZX-D onwards), Centaur(ZX-A onwards)

WuDaoKou isn't derivative of Isaiah, but rather a re-engineered/re-designed architecture; increased decode/improved pipeline depth/re-organized scheduler&reservation ports&latency/etc. Much like Bobcat(Fam 14h) vs Jaguar(Fam 16h) or in this case Zen(Fam 17h) and Dhyana(Fam 18h).

ZX-D/ZX-E/ZX-F/ZX-G are all 100% homegrown. (Primarily to get the processor certificates for allowance into (mission-critical?) government systems)
ZX-A/ZX-B/ZX-C aren't 100% homegrown. (ZX-C 100% being a relabeled VIA chip using Centaur's CNR cores, like the products before it)
 
Last edited:

NostaSeronx

Diamond Member
Sep 18, 2011
3,686
1,221
136
KX-6000 series is still derivative of Isaiah. Zhaoxin needed a clean-sheet design some time ago.
If that was the case, what happened to AVX2?
CNR has AVX2, WDK/LJZ don't.

Absolute clean sheet(massive improvement of perf) was always going to be ZX-F on sub-7nm.

I do not believe after all they have achieved they will return to daddy Centaur and ask for CNS and AVX512. Instead, they are going to build their own core and their own instruction set extension. They aren't going to self-ban themselves again from the lucrative government deals.
 
Last edited:

Kosusko

Member
Nov 10, 2019
161
120
116
VIA x86 CPU = Centaur
Zhaoxin x86 CPU = Zhaoxin(ZX-D onwards), Centaur(ZX-A onwards)

WuDaoKou isn't derivative of Isaiah, but rather a re-engineered/re-designed architecture; increased decode/improved pipeline depth/re-organized scheduler&reservation ports&latency/etc. Much like Bobcat(Fam 14h) vs Jaguar(Fam 16h) or in this case Zen(Fam 17h) and Dhyana(Fam 18h).

ZX-D/ZX-E/ZX-F/ZX-G are all 100% homegrown. (Primarily to get the processor certificates for allowance into (mission-critical?) government systems)
ZX-A/ZX-B/ZX-C aren't 100% homegrown. (ZX-C 100% being a relabeled VIA chip using Centaur's CNR cores, like the products before it)

90mn VIA C7/C7-M/C7-D/Eden (C5J) = CentaurHauls Family 6 Model 13 Stepping 0

-----------------------------

65mn VIA Nano L2000/U2000 Series (CNA) = CentaurHauls Family 6 Model 15 Stepping 2

65mn VIA Nano L3000/U3000 Series (CNB) = CentaurHauls Family 6 Model 15 Stepping 10
40mn VIA Nano X2/EdenX2 (CNC), QuadCore L4700/U4650 (CNQ) = CentaurHauls Family 6 Model 15 Stepping 13

28nm VIA EdenX4 C4000 Series / QuadCore C4000 Series (CNR): CentaurHauls Family 6 Model 15 Stepping 14
28nm C4000 Series (ZX-C+) = CentaurHauls Family 6 Model 15 Stepping 14

-----------------------------

28nm KX-5000 Series (ZX-D) = CentaurHauls Family 7 Model 11 Stepping 1
16nm KX-6000 Series (ZX-E) = CentaurHauls Family 7 Model 11 Stepping 1

-----------------------------

16nm Centaur CHA (CNS - ZX-F) = CentaurHauls Family 6 Model 71 Stepping 1
 
Last edited: