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

Page 15 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
Gracemont really upped Intel's game performance-wise, though efficiency may have taken a back seat. VIA/Zhaoxin really can't compete with Gracemont, and tbh they weren't competitive with Tremont either. Intel just chose not to sell Tremont to many customers in core configurations above 4c (and when they did, it was in various network appliances etc.).

As for the merger, we have another thread for that so I'll take myself over there!

No wonder car makers cannot find silicon, when stuff like this soaks up global production capacity.

Ehhh

Auto mfgs are often using ancient nodes from 10+ years ago. I think you misapprehend some part of that situation. And Zhaoxin's volumes are pretty low.
 

Jwilliams01207

Junior Member
Dec 6, 2013
24
2
71
I think I know what's holding up Zhaoxin from releasing new hardware. TSMC won't manufacture for them anymore so I guess their only option is SMIC. We can expect to see a CPU release from them in 2021 which will coincide with the availability of SMIC's new N+1 process ...

5nm will be tricky for SMIC since they can't acquire any EUV scanner but I guess we'll see a domestic solution in the next 5 years ...
 

NTMBK

Lifer
Nov 14, 2011
10,207
4,939
136
I think I know what's holding up Zhaoxin from releasing new hardware. TSMC won't manufacture for them anymore so I guess their only option is SMIC. We can expect to see a CPU release from them in 2021 which will coincide with the availability of SMIC's new N+1 process ...

5nm will be tricky for SMIC since they can't acquire any EUV scanner but I guess we'll see a domestic solution in the next 5 years ...

Where did this meme come from? There's no evidence that TSMC are refusing to work with Zhaoxin.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
TSMC isn't a part of the anti-China embargo of which I'm aware. Though the Huawei situation made it look that way.
 

Kosusko

Member
Nov 10, 2019
161
119
116
The King Is DeadLong Live the Queen!
New x86 processors from ZHAOXIN @ 2022

CgAG0mHUCViAM3L8AADOm06re3M859.jpg

source: https://www.zhaoxin.com/news_view.aspx?nid=2&typeid=267&id=1397


P.S. Goodbye Centaur Technology ;)
 

SteinFG

Senior member
Dec 29, 2021
395
450
106
That's straight up i7 6950x, same ihs, same die shape, wtf
edit: i'm dumb, that's two different processoors in the pic
 
Last edited:

Kosusko

Member
Nov 10, 2019
161
119
116
Processor shown on the left is Kaisheng KH-40000 Series, I hope with CNS design and on the right I'd guess Kaixian KX-6000G Series with new integrated graphics processor.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Here's hoping they actually launch something competitive this year!
Don't even hope for it.
CNS/CHA barely beats custom FTC663 cores.

Zhaoxin is also only implementing their CNS/CHA cores in quad-core, so that multithreading lead is completely gone.

In-house x86 in China is pretty much dead w/ custom ARM & RISC-V designs.
Zhaoxin next-gen server = 16-cores
ARM & RISC-V next-gen server in China = 64-128 cores
 
Last edited:

Kosusko

Member
Nov 10, 2019
161
119
116
How are the dual dies connected?
I think with new ZPI "3.0" technology by new Kaisheng KH-40000 Series.

The Kaisheng KH-40000 series of processors are expected to be officially launched in 2021 (now 2022). The product has a new independent CPU micro-architecture design, based on 16nm technology, and continues to use the SoC solution. The number of CPU cores for a single processor reaches the existing Kaisheng KH- 4 times that of 30000 series processors. At the same time, KH-40000 will continue to support dual-channel interconnection, that is, up to 64 cores in the system, and support DDR4 memory and PCIe 3.0.
source: https://www.zhaoxin.com/news_view.aspx?nid=2&typeid=201&id=827
__________________________________________________________________________________________________________________
Kaisheng KH-30000 series processors also support the new inter-chip interconnection ZPI 2.0 technology and support point-to-point multi-channel processor expansion. ZPI (Zhaoxin Processor Interconnect) is an inter-chip interconnection technology independently developed by Zhaoxin. The KH-30000 series processors support dual-channel interconnection, that is, it supports the integration of 2 processors into a ccNUMA (Cache Coherent Non-Uniform Memory Access) system, which can support up to 16 cores, bringing double performance to the system.
CgAH517i5y6APMlCAAA2OAeQKDU281.jpg

source: https://www.zhaoxin.com/NewsStd_1049.html
source: http://blog.jc170.com/?id=1099

__________________________________________________________________________________________________________________
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Zhaoxin next-gen server = 16-cores
Dual Die = 32-cores
Dual Socket = 64-cores
Nope, it is 16-cores with two KX-7000 dies and two sockets...
The Kaisheng KH-40000 series of processors are expected to be officially launched in 2021 (now 2022). The product has a new independent CPU micro-architecture design, based on 16nm technology, and continues to use the SoC solution. The number of CPU cores for a single processor reaches the existing Kaisheng KH- 4 times that of 30000 series processors. At the same time, KH-40000 will continue to support dual-channel interconnection, that is, up to 64 cores in the system, and support DDR4 memory and PCIe 3.0.
source: https://www.zhaoxin.com/news_view.aspx?nid=2&typeid=201&id=827
That information is out of date, since they lost access to 7nm. Only through 7nm would they have gotten 4x the core count.
zhaoxin.jpg
16 cores with dual-die/dual-socket
zhaoxin2.jpg

Done like this:
zhaoxin3.jpg
Gives the 8DDR4, and up to 176 PCIe, of which 48 is interconnected through the off-die interconnect chip.
 
Last edited:

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,439
14,408
136
I really don't understand why you keep posting about these. They are much slower than anything else out there.

Are you are marketing salesman for them ?
 
  • Love
Reactions: Thunder 57

The Hardcard

Member
Oct 19, 2021
46
38
51
I really don't understand why you keep posting about these. They are much slower than anything else out there.

Are you are marketing salesman for them ?

I have an interest in all attempts at advanced technology, regardless of capabilities, resources, or politics.

Is this slower than Elbius? I would read an article pitting this and Elbius against Excavator, Haswell, Zen 3, Alder Lake, 8cx Gen 1, and M1. I would love to see where they land on a range of code?

Separately, I am also interested in Power 10, especially with memory inception examples, though that would require IBM or one of their vendors to be really generous to a reviewer.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
@The Hardcard

If you look further back in this thread, you can find a few benchmarks of the current and previous generations of VIA/Zhaoxin CPUs. No idea how that stacks up against the current Elbrus chips since they're so hard to get in the West, generally don't run the same software, and . . . Bob's your uncle.
 

The Hardcard

Member
Oct 19, 2021
46
38
51
@The Hardcard

If you look further back in this thread, you can find a few benchmarks of the current and previous generations of VIA/Zhaoxin CPUs. No idea how that stacks up against the current Elbrus chips since they're so hard to get in the West, generally don't run the same software, and . . . Bob's your uncle.

Oh yeah, I been following these threads. I was just letting Markfw know that interest remains even after this latest engineering team wierdness.

With all the money flowing into competing silicon, I believe there will eventually again be money behind a leading edge team using this x86 license. I’m baffled as to why it hasn’t yet happened, th has to be some kind of madness somwhere in this company.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,439
14,408
136
I have an interest in all attempts at advanced technology, regardless of capabilities, resources, or politics.

Is this slower than Elbius? I would read an article pitting this and Elbius against Excavator, Haswell, Zen 3, Alder Lake, 8cx Gen 1, and M1. I would love to see where they land on a range of code?

Separately, I am also interested in Power 10, especially with memory inception examples, though that would require IBM or one of their vendors to be really generous to a reviewer.
I thought I quote Kosusko, who I was replying to. This was not directed at you.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
I’m baffled as to why it hasn’t yet happened, th has to be some kind of madness somwhere in this company.

At least up til this point, Centaur-related products have not been performant vs the rest of the market in decades. Something to think about. They used to have some innovation in the low-power segment, but Atom et al killed that ages ago. Today, the CNS/KX line of CPUs burns too much power to be taken seriously at its performance level. That's not really going to attract much investment. You'd have to bring in your own team of crack engineers to build something from the ground-up AND overcome the stigma of past poor performance of VIA/Zhaoxin CPUs. It would be a tall order. It's easier to just clone reference ARM CPUs and cheaper to dabble in RISC-V.

Now that Intel has hired off Zhaoxin's core engineering team it's not really clear what is going to happen.
 

The Hardcard

Member
Oct 19, 2021
46
38
51
At least up til this point, Centaur-related products have not been performant vs the rest of the market in decades. Something to think about. They used to have some innovation in the low-power segment, but Atom et al killed that ages ago. Today, the CNS/KX line of CPUs burns too much power to be taken seriously at its performance level. That's not really going to attract much investment. You'd have to bring in your own team of crack engineers to build something from the ground-up AND overcome the stigma of past poor performance of VIA/Zhaoxin CPUs. It would be a tall order. It's easier to just clone reference ARM CPUs and cheaper to dabble in RISC-V.

Now that Intel has hired off Zhaoxin's core engineering team it's not really clear what is going to happen.

I was thinking more of investing in making performant, competitive products with this x86 license. If you are going spend huge sums to employ a competent semiconductor design team and then pay the sky high costs to have them create designs on advanced nodes, having a processor that only requires current code to be optimized rather than rewritten seems the a likely quicker path to return on investment.

Theoretically, it seems like someone could bankroll VIA to make marketable chips for some negotiated cut of the return, but there’s apparently many devils in the details.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
I was thinking more of investing in making performant, competitive products with this x86 license.

I addressed that directly. Zhaoxin more-or-less did what you suggested, though it's unclear how much money has been committed to their effort to date. What is clear is that they have not succeeded in producing a competitive CPU. There is the very real risk that anyone who attempted to buy out Zhaoxin would meet the same fate, no matter how many financial resources they committed to their new company. The inherent value of owning an x86 license today isn't so great that spending billions on a design team, facilities, and wafer commitments from a major fab like Samsung (you would be on a waiting list at TSMC) makes any sense.
 

The Hardcard

Member
Oct 19, 2021
46
38
51
I addressed that directly. Zhaoxin more-or-less did what you suggested, though it's unclear how much money has been committed to their effort to date. What is clear is that they have not succeeded in producing a competitive CPU. There is the very real risk that anyone who attempted to buy out Zhaoxin would meet the same fate, no matter how many financial resources they committed to their new company. The inherent value of owning an x86 license today isn't so great that spending billions on a design team, facilities, and wafer commitments from a major fab like Samsung (you would be on a waiting list at TSMC) makes any sense.

All investments involve the same or greater risk. You just described precisely the risk AMD investors were required to take in financing Zen development.

ARM and RISC-V are bigger risks precisely because it requires customers to move away from codebases they are familiar with and have whatever level of confidence in to switch to new codebases. Enterprise customers for these have to - in addition to the hardware buy - drop dollars on new software development and maintenance.

From a purely technological versus getting the enterprise to adopt said technologies view, I’d say Ampere, Marvell, Qualcomm/Nuvia and other are taking far bigger risks. There appear to be other entities that want in on the action given that there are several no name companies buying 7 and 5nm wafers from TSMC and Samsung, as well as dozens putting their designs on 14/12 nm, which is only cheap relatively.

Does it make sense to take these extremely costly risks? I don’t know, but it is happening. Not only are ARM spending money on new architectures with no guarantee of success, but there is growing interest in spending billions more for custom architectures based on ARM. Any hope of a modern performance level RISC-V involves spending money on a new advanced node architecture with a steeper climb to success than an x86 design.

Zhaoxin is in some mystery status, but the VIA license remains in a climate where multiple groups of people are willing to spend billions on new designs that are all moonshots or crapshoots with no way to tell which is which. I think if VIA ever got serious about acquiring the people and resources to mix it up, plenty of money people would take that chance.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
All investments involve the same or greater risk. You just described precisely the risk AMD investors were required to take in financing Zen development.

AMD nearly tanked, despite having several years of performance dominance prior to Q2/Q3 2006. The only reason why AMD got away with Zen is that they already had a focused team of engineers willing to take another shot at the higher echelons of the x86 market. And they hedged their bets by simultaneously developing an ARM core, which they later shelved when it became clear that x86 Zen would be a winner.

If in 2012-2013 an investor had taken the same amount of cash and the same debt profile as what AMD faced at the time and attempted to assemble a team of engineers to produce a competitor to Zen (and Intel), they would have failed. 100% guaranteed. That's too much intertia to overcome.

ARM and RISC-V are bigger risks

ARM? Certainly not. You can hire a much smaller team of engineers and become the next Rockchip by licensing designs directly from ARM, Ltd. Long term you wind up paying licensing fees which may cut into profits, but short term, it's much cheaper to get started that way. RISC-V does not (yet) have a good library of reference designs, but then you aren't paying any licensing fees either.

precisely because it requires customers to move away from codebases

In the consumer realm, this is a non-issue. Not many people want to use old software anymore, except maybe for games. In server maybe, but ARM server is better-represented in the software realm now. RISC-V may follow if anyone chooses to take the uarch family seriously-enough to produce performant server/workstation parts.

From a purely technological versus getting the enterprise to adopt said technologies view, I’d say Ampere, Marvell, Qualcomm/Nuvia and other are taking far bigger risks.

Only because Amazon, MS, Google, etc. have chosen to hire in-house design teams so that they don't have to buy ARM solutions from other parties. Nuvia got their payout and Ampere is still kicking.

In conclusion: it would take hundreds of millions of dollars just to acquire Zhaoxin's x86 license, and even more money to get what remains of their engineering talent. It also isn't clear that the Centaur license is transferrable, e.g. a complete transfer of the license may terminate it. Still not 100% sure how Zhaoxin has avoided that problem, but I think it's because technically they're still in a merger/cooperative agreement situation with VIA/Centaur. Regardless, all that spending plus billions more for a clean-sheet design will likely result in years of poor performance relative to the rest of the x86 market. The ARM world will also be leaving you in the dust.