Interesting results. I wish we knew more about what node it’s on. That sound like maybe 22nm level efficiency. It’s still seemingly much more performant than the strongest RISC V cores to come out of China so far.
Also VIA Next advertised 6nm project on their website.
Last year VIA taped out 5nm project.
It's propably VIA, who send Zhaoxin projects to production at TSMC. Like AMD-Hygon, what is on entity list.
I'll be not suprised if higher clocked kx-7000 is from TSMC and lower clocked versions are produced on 7nm SMIC. IOD is slow, and GPU c-1190 on IOD has 2/3 clock of c-1080, so maybe it is 16/12 SMIC.
Also VIA Next advertised 6nm project on their website.
Last year VIA taped out 5nm project.
It's propably VIA, who send Zhaoxin projects to production at TSMC. Like AMD-Hygon, what is on entity list.
I'll be not suprised if higher clocked kx-7000 is from TSMC and lower clocked versions are produced on 7nm SMIC. IOD is slow, and GPU c-1190 on IOD has 2/3 clock of c-1080, so maybe it is 16/12 SMIC.
How does it compare in energy use? Does it consume less or more for a give task, compared to newer AMD or Intel, or even Bulldozer-era, chips? Thank you for the write-up.
@Chester: Thanks! I would love to see more of these deep dives into CPUs that we here in the US usually don't see or even know much about!
A few comments/observations:
- I see this PRC-homegrown x64 CPU more as an insurance policy by their government than anything else. In addition to full control over the microcode and firmware (security), a lot of important software is still x86/x64 native. Having their fully homegrown x64 CPUs also shields China against getting a possible complete cut off from getting new AMD and Intel CPUs.
- The historically poorer performance of the L3 Cache in Intel's CPUs (vs. AMD since Ryzen) is still with us, and continues to be a bit of an Achilles heel for Intel. The superiority of AMDs X3D Ryzens in gaming and certain productivity apps is not simply due to just having more L3 cache. L3 caches of modern Ryzens also have significantly lower latencies and better throughput than their Intel counterparts. I wonder if you or George ever had a chance to speak to one of Intel's lead CPU architects about this.
I think the large L2 on recent Intel chips like Arrow Lake goes a long way toward mitigating higher L3 latency. For perspective 3 MB is like the last level cache on some 2c/4t parts from 12 years ago.
But with the KX-7000, they have 512 KB of L2 and even worse L3 latency. Not a great combo.
Ryzen has also always showed significant advantage in L3 bandwidth. That’s why they don’t lose near as much performance as Intel spilling from L2 to L3
Can you write a blog comparing Chinese domestic CPUs, including Loongson’s 3A6000, Zhaoxin’s KX-7000, and Phytium’s D3000? Many people are curious about which of these companies has the most potential.
Hi all, I have motherboad ASUS XC-KX700M D4 with this CPU KX-7000. In my test power consumpting whole system is like below.
KX-7000 @3GHz turbo off with iGPU + 2x8GB DDR4 3200MHz + Nvme 128GB:
- Idle: 32W
- Cinebench R23 multi: 100W
KX-7000 @3-3,6GHz turbo on with iGPU + 2x8GB DDR4 3200MHz + Nvme 128GB:
- Idle: 32W
- Cinebench R23 multi: 120W
Interesting results. I wish we knew more about what node it’s on. That sound like maybe 22nm level efficiency. It’s still seemingly much more performant than the strongest RISC V cores to come out of China so far.
KX-7000 @3GHz turbo off: Cinebench R23 multi 4469 - 105W
KX-7000 @3-3,6GHz turbo on: Cinebench R23 multi 4725 - 125W
Ryzen 7 1700 @3GHz turbo off SMT off: Cinebench R23 multi 6082 90W
Ryzen 7 1700 @3GHz turbo off SMT on: Cinebench R23 multi 7706 105W
Total power whole platforms: CPU+board+2x8gb ddr4+Nvidia T400+nvme 128gb+psu
Interesting! What were the benchmark numbers you were able to get?
My results.
KX-7000 @3GHz turbo off
- Cinebench R23 single: 582
- Cinebench R23 multi: 4469
- CPU-Z single: 315
- CPU-Z multi: 2540
KX-7000 @3-3,6GHz turbo on
- Cinebench R23 single: 605
- Cinebench R23 multi: 4725
- CPU-Z single: 350
- CPU-Z multi: 2705
For comparision:
Ryzen 7 1700 @3GHz turbo off SMT off
- Cinebench R23 single: 786
- Cinebench R23 multi: 6082
- CPU-Z single: 345
- CPU-Z multi: 2770
Ryzen 7 1700 @3GHz turbo off SMT on
- Cinebench R23 single: 786
- Cinebench R23 multi: 7706
- CPU-Z single: 345
- CPU-Z multi: 3760
In my tests IPC of KX-7000 is between Excavator and Zen1.
In some test (AIDA64 for example) it can reach level Zen2, Zen3 due to having FPU 256bit.
More test in my comments in forum anantech:
https://forums.anandtech.com/threads/zhaoxins-zx-f-kx-7000-kh-40000-and-beyond.2564975/post-41432844
https://forums.anandtech.com/threads/zhaoxins-zx-f-kx-7000-kh-40000-and-beyond.2564975/post-41433486
I get better figures with a 5600u @ 15w
You’re also comparing a modern CPU with up to date architecture on 7nm to a smallish company working with probably 22nm at best.
Zen 2 was built on 7nm. Ryzen 7 1700 is Zen 1, built on 14nm at GlobalFoundries.
Zen+ used 12nm GF (tweaked 14nm), Zen 2 (2019) used 7nm at TSMC.
5600u is Zen3, I guess.
Thanks, also for the links to your posts with additional results!
Maybe I missed that in Chester's write-up, but I didn't realize that the Zhaoxin CPU uses the LGA1700 socket. Did they in-license that IP from Intel?
It’s 6nm for CCD. https://forums.anandtech.com/threads/zhaoxins-zx-f-kx-7000-kh-40000-and-beyond.2564975/page-20#post-41133323
Also VIA Next advertised 6nm project on their website.
Last year VIA taped out 5nm project.
It's propably VIA, who send Zhaoxin projects to production at TSMC. Like AMD-Hygon, what is on entity list.
I'll be not suprised if higher clocked kx-7000 is from TSMC and lower clocked versions are produced on 7nm SMIC. IOD is slow, and GPU c-1190 on IOD has 2/3 clock of c-1080, so maybe it is 16/12 SMIC.
I'm interested in these tests, is there a corresponding github to reproduce the results?
There is, but I want to encourage others to write their own tests. My github is really a mess of experiments rather than something presentable
It’s 6nm for CCD. https://forums.anandtech.com/threads/zhaoxins-zx-f-kx-7000-kh-40000-and-beyond.2564975/page-20#post-41133323
Also VIA Next advertised 6nm project on their website.
Last year VIA taped out 5nm project.
It's propably VIA, who send Zhaoxin projects to production at TSMC. Like AMD-Hygon, what is on entity list.
I'll be not suprised if higher clocked kx-7000 is from TSMC and lower clocked versions are produced on 7nm SMIC. IOD is slow, and GPU c-1190 on IOD has 2/3 clock of c-1080, so maybe it is 16/12 SMIC.
How does it compare in energy use? Does it consume less or more for a give task, compared to newer AMD or Intel, or even Bulldozer-era, chips? Thank you for the write-up.
@Chester: Thanks! I would love to see more of these deep dives into CPUs that we here in the US usually don't see or even know much about!
A few comments/observations:
- I see this PRC-homegrown x64 CPU more as an insurance policy by their government than anything else. In addition to full control over the microcode and firmware (security), a lot of important software is still x86/x64 native. Having their fully homegrown x64 CPUs also shields China against getting a possible complete cut off from getting new AMD and Intel CPUs.
- The historically poorer performance of the L3 Cache in Intel's CPUs (vs. AMD since Ryzen) is still with us, and continues to be a bit of an Achilles heel for Intel. The superiority of AMDs X3D Ryzens in gaming and certain productivity apps is not simply due to just having more L3 cache. L3 caches of modern Ryzens also have significantly lower latencies and better throughput than their Intel counterparts. I wonder if you or George ever had a chance to speak to one of Intel's lead CPU architects about this.
I think the large L2 on recent Intel chips like Arrow Lake goes a long way toward mitigating higher L3 latency. For perspective 3 MB is like the last level cache on some 2c/4t parts from 12 years ago.
But with the KX-7000, they have 512 KB of L2 and even worse L3 latency. Not a great combo.
Yes, in this aspect, the KX-7000 got the worst of both worlds; small-ish L2 and high latency L3.
Ryzen has also always showed significant advantage in L3 bandwidth. That’s why they don’t lose near as much performance as Intel spilling from L2 to L3
Can you write a blog comparing Chinese domestic CPUs, including Loongson’s 3A6000, Zhaoxin’s KX-7000, and Phytium’s D3000? Many people are curious about which of these companies has the most potential.
Which one do you think is better between the 3A6000 and the KX7000? Both of which are made in China.
The 3A6000 offers better IPC, but it suffers from serious software issues. The x86, on the other hand, should have far fewer compatibility problems.