Zhaoxin's KX-7000 8-Core Processor Tested in Detail, Bested by 7 Year Old Core i3

[Fouquin]

PC Watch recently got hands-on with Shanghai Zhaoxin's latest desktop processor for some in depth testing and published a less than optimistic review comparing it to both the previous generation KX-U6780A and Intel's equally clocked budget quad-core offering from 2017, the 3.6 GHz Core i3-8100. Though Zhaoxin's latest could muscle its way through some multithreaded tests such as Cinebench R23 due to having twice the core count, the single core performance showed to be nearly half that of the i3 in everything from synthetic tests to gaming.

PC Watch tested with the Dragon Quest X Benchmark, a DX9.0c title, to put the spotlight on single core gaming performance even in older games as well as with Final Fantasy XIV running the latest Golden Legacy benchmark released back in April of this year to show off more modern multithreaded gaming. With AMD's RX 6400 handling graphics at 1080p the KX-7000/8 scored around 60% of the i3-8100 in Dragon Quest X, and in Final Fantasy XIV it scored 90% of the i3. The result in Final Fantasy XIV was considered, "somewhat comfortable" for gameplay but still less than optimal. As a comparison point for a modern budget gaming PC option the Ryzen 5 5600G was also included in testing, where in Final Fantasy XIV it was 30% ahead of the KX-7000/8. PC Watch attempted to put the integrated ZX-C1190 to work in games but found that despite supporting modern APIs and features, the performance was no match for the competition.



Zhaoxin originally announced KX-7000 to be built on a 7 nm FinFET process and based on their own "Century Avenue" core architecture (presumed to be a modified variant of Centaur Technologies cancelled "CNS" core architecture) back in 2019, with rumors surrounding a 2021 release target. This date would shift up through the years until the chips finally launched in China in December 2023. Originally released in OEM machines, the KX-7000 appears to finally be receiving support from DIY vendors, with ASUS seemingly producing the mATX motherboard that PC Watch used in their review. With KX-7000 Zhaoxin has moved away from the previous generation's BGA packaging and into Intel's LGA1700 socket. Though electrically incompatible, this move may ease its adoption to the DIY market.

You can check out PC Watch's full review here.

View at TechPowerUp Main Site | Source

 

[Frank_100]

To be fair, 2017 Intel cpu's were very good.
Lots of headroom.

 

[Ruru]

To be fair, 2017 Intel cpu's were very good.
Lots of headroom.

i3-8100 is practically an i5-7500 or similar. While those old 4c/4t Intels may not be gaming beasts anymore, they're still more than usable in basic usage.

 

[DirtyDingusMcgee]

China always has and always will make the best processors. There are no processors superior to Zhaoxin or Kirin.

(Hopefully this makes up for those nasty things I said about Huawei......come on rising social credit score).

 

[csendesmark]

Can someone tell me what kind of CPU is this?
Like: Is this x86?
I mean, isn't x86 a closed license? You need AMD Intel and whomever owns Cyrix now to agree to let other companies use it?

 

[regs]

This isn't CNS. This is still evolution of Nano.

Can someone tell me what kind of CPU is this?
Like: Is this x86?
I mean, isn't x86 a closed license? You need AMD Intel and whomever owns Cyrix now to agree to let other companies use it?

Continuation of Centaur's WinChip and Cyrix.

With KX-7000 Zhaoxin has moved away from the previous generation's BGA packaging and into Intel's LGA1700 socket.

They didn't move from BGA. This is just LGA version. So as previous generation had LGA and BGA versions.

开先® KX-7000系列处理器 - 兆芯

www.zhaoxin.com

 

[Fouquin]

This isn't CNS. This is still evolution of Nano.

Evolution of Isaiah/Nano mostly died with ZX-E when they reordered the cache subsystem and dropped from 3-way to 2-way decode logic, as well as dropping AVX2 from Isaiah II/ZX-C+. ZX-F would appear to share more with CNS than it does Isaiah in design language, whether it directly lifted that IP or not is yet to be proven but what is known is CNS is still VIA's IP and Zhaoxin would have access to license portions of it.

The cache arrangement alone is dissimilar from any Nano/CNQ/CNR base architecture. Per-core L2 caches (as opposed to a module-shared L2 or module-private L2 such as those with CNQ/CNR) with a global 16-way associative L3 cache is very similar to CNS's cache configuration, with L2 and L3 being doubled in capacity and L2 being limited to 8-way associative versus 16. KX-7000 are all still family 7 (though ZX-C+ and CNS are both family 6 despite sharing very little underlying groundwork) and performance numbers suggest it is still a very slim core, but they definitely took some parts of CNS into consideration when they laid down KX-7000.

The Geekbench result (I know, groan, but it provides a 1:1 comparison here) tells a lot of the story. KX-7000 is no 2-wide decode/7-port execution core. This has significant improvements beyond just a clock increase on a new node. But I grant you, it is not a direct copy of CNS. It is however not far off when you make the necessary clock offset adjustment (1.4x in this case, the CNS was run at 2415 MHz not the 2300 MHz shown.)

They didn't move from BGA. This is just LGA version. So as previous generation had LGA and BGA versions.

Only their server chips were on LGA prior. On the Zhaoxin page you just linked it shows all prior consumer variants as HFCBGA or FCBGA. KX-7000 is the first generation to offer LGA on the consumer part to a DIY audience.

 

[Gigaherz]

Can you overclock these?

 

[Crackong]

Fueled by patriotism
Funded by patriotism
Bought by patriotism

No one wants them outside of patriotism.

 

[kondamin]

i don’t get why they haven't dropped the isa they are hopelessly behind intel And amd i think loongson has surpassed them now

 

[MacZ]

This kind of CPU is enough to run 80% of an economy.

You don't need a self-destructing 14900K to use word processors and spreadsheet programs.

The gaming angle is irrelevant, as games are mostly a time and energy sink.

 

[kondamin]

This kind of CPU is enough to run 80% of an economy.

You don't need a self-destructing 14900K to use word processors and spreadsheet programs.

The gaming angle is irrelevant, as games are mostly a time and energy sink.

100% of an economy if you go back far enough.
i disagree about the gaming, gaming is probably the main reason it has become as pervasive as it is.

 

[tpuuser256]

Being able to make a decent CPU in-house is a huge achievement, even though they probably turbo copied the west and hired western engineer to soulsuck their skills

 

[Wirko]

The reviewers should have also made an X-ray image of the processor, so we'd know the die sizes. Then someone could make an informed guess about the process node for each chiplet.

 

[raising19]

There is a problem with the BIOS of this motherboard, involving memory bandwidth and frequency

 

[close]

While it sounds bad at first, that level of performance is quite decent (I'm still daily on 6600U and 6700 and they're more than adequate even today). And it still beats what most of the world has built (looking at EU having nothing close to this). For a country like China technological independence is probably more important than cutting edge performance.

 

[AnotherReader]

In many ways, this is a regression compared to Centaur's CNS. Chips and Cheese examined one and found it to be slightly behind equivalently clocked Haswell in single threaded applications. CNS supported AVX-512, but this core doesn't.

Note how AVX-512 allows the CNS to surpass Haswell at the same clocks.

 

[regs]

Only their server chips were on LGA prior. On the Zhaoxin page you just linked it shows all prior consumer variants as HFCBGA or FCBGA. KX-7000 is the first generation to offer LGA on the consumer part to a DIY audience.

Yes, KH-40000, which are previous generation KX-E.

The cache arrangement alone is dissimilar from any Nano/CNQ/CNR base architecture. Per-core L2 caches (as opposed to a module-shared L2 or module-private L2 such as those with CNQ/CNR) with a global 16-way associative L3 cache is very similar to CNS's cache configuration, with L2 and L3 being doubled in capacity and L2 being limited to 8-way associative versus 16.

Doesn't really tell much. They have a tech that they keep improving for over decade. It's probably had some ideas, but it's still not a performance core. That one is yet to come.

The Geekbench result (I know, groan, but it provides a 1:1 comparison here) tells a lot of the story. KX-7000 is no 2-wide decode/7-port execution core.

There is more interesting comparison with CHA itself. Mind CHA is running at 2.5 GHz, which ZX-F is boosting at 3.4 GHz.

KaiTian 90WVA009KX vs Centaur Technology CHA001 MB - Geekbench

 

[Wirko]

The L3 cache, if it's actually 32 MB, is also disproportionately large for the performance. Intel's 8-core 14 nm processors had half as much.

 

[AnotherReader]

Yes, KH-40000, which are previous generation KX-E.


Doesn't really tell much. They have a tech that they keep improving for over decade. It's probably had some ideas, but it's still not a performance core. That one is yet to come.


There is more interesting comparison with CHA itself. Mind CHA is running at 2.5 GHz, which ZX-F is boosting at 3.4 GHz.

KaiTian 90WVA009KX vs Centaur Technology CHA001 MB - Geekbench

It seems that this is a significant regression in IPC compared to Centaur's last hurrah.

 

[Fouquin]

Doesn't really tell much. They have a tech that they keep improving for over decade. It's probably had some ideas, but it's still not a performance core. That one is yet to come.

They don't have over a decade of improvements, though. Zhaoxin was founded in 2013 and simply shipped licensed VIA/Centaur chips until the release of ZX-D in 2017, which took them an estimated two years to develop. ZX-D combined ZX-C+'s quad-core modules and integrated NB/IMC but kept the core and cache configurations nearly identical, so while it was a new design it was not a new architecture. Their first proper core redesign came with ZX-E in 2018 and that was simply ZX-D cut down by 50% with an improved module-shared L2 cache and iGPU, so still licensing the architecture IP from VIA. If we're generous we say they spent 3 years on taking an existing design and cutting it down to turn into an SoC. Around 2019 they would announce they were beginning development on ZX-F/KX-7000 for 7nm but it would remain suspiciously absent from their 2020 roadmap. So in total that gives them 8 years of new design experience and only two full designs from their own teams. Yes, Wang Weilin has been working in CPU design for two decades, but his teams and especially Zhaoxin have not been creating ground-up designs for that long.

CNS was up and running in Austin, TX in 2019 when Zhaoxin announced they were designing KX-7000. The CHA SoC was complete and A1 stepping chips were packaged and in testing. A2 and A3 would be ready by the middle of 2020 with the intent to ship before the end of the year. As previously stated, Zhaoxin showed their roadmap in 2020 and KX-7000 was entirely unrepresented. That's strange, because in the lead-up to KX-6000 Zhaoxin already had the core count and frequency targets locked in. Many would suspect this is because they were waiting for the October 2020 IP deals which included a $138M purchase of x86 chip design IP from VIA's subsidiary VIABASE as well as the $118M purchase from VIATECH that held "processor-related technology, materials and other intellectual property rights". Strangely around the same time that Centaur's CHA was supposed to reach availability...

So in 4 years we've seen Zhaoxin's "small" core double in performance and come within 20% of what CNS was accomplishing in 2019 right before Zhaoxin was sold a buttload of VIA/Centaur x86 design IP, and somehow that doesn't suggest that parts of CNS's core design isn't in KX-7000? They aren't identical, I'll give you that. KX-7000 is still a slimmed down core from CNS, but to say they the two have no relation despite them both showing equal strengths in comparison to ZX-E? And with the timeline of events showing that Zhaoxin went from having access to only older IP to suddenly being flush with ~$260M worth of brand new IP, having no stated specifications for KX-7000 in 2020 to shipping it as a "brand new core architecture" in 2023? All signs point to KX-7000 being related to CNS. Yes, Zhaoxin does still have a "Bigger" core in the works. It too is likely leveraging that same IP and will be a frequency increased CNS with an improved cache and China's SM2/SM3/SM4 PadLock features.

There is more interesting comparison with CHA itself. Mind CHA is running at 2.5 GHz, which ZX-F is boosting at 3.4 GHz.
KaiTian 90WVA009KX vs Centaur Technology CHA001 MB - Geekbench

Which I also linked to. The results shown support my position on it using parts of CNS IP. The improvements shown over ZX-E are not generational, they are exponential. This is no, "we improved the core 50% gen-over-gen." This is 210% faster than ZX-E in a single generation, clock normalized. What else is over 200% faster than ZX-E per clock? Centaur's CNS.

 

[psydroid]

i don’t get why they haven't dropped the isa they are hopelessly behind intel And amd i think loongson has surpassed them now

Loongson is indeed China's flagship computer chip. Loongson has always been ahead of Zhaoxin and is now even rivaling recent designs from Intel and AMD.

Zhaoxin is just a fallback for running legacy x86 applications and games natively.

 

[KLMR]

How long to take over western designs? That should be the point. Now that Intel and AMD may have a competitor they seem to crumble appart.
The fact that China and not Europe decided to battle the cpu market is also symptomatic...

 

[AusWolf]

How long to take over western designs? That should be the point. Now that Intel and AMD may have a competitor they seem to crumble appart.
The fact that China and not Europe decided to battle the cpu market is also symptomatic...

Why would Europe want to compete? We're nothing more but the lapdogs of our American masters (politically at least).

 

[AnotherReader]

Loongson is indeed China's flagship computer chip. Loongson has always been ahead of Zhaoxin and is now even rivaling recent designs from Intel and AMD.

Zhaoxin is just a fallback for running legacy x86 applications and games natively.

Loongson is much better than Zhaoxin, but it is a long way behind AMD and Intel.