News The only Chinese chipmaker with an x86 license releases surprisingly modern new chips made with a mystery process node —Xhaoxin KX-7000 CPU launche...

[danyuzunov]

Ah ye, the natively made chinese CPU made with the TSMC 7nm process

 

[Albert.Thomas]

Eh, that's not what they say.

Here are some details about a prior, 28 nm SoC they developed 6 years ago.

​

"WuDaoKou is largely a brand new architecture designed by Zhaoxin. This is a departure from earlier microarchitectures such as ZhangJiang which were a lightly modified version of VIA Technologies (Centaur) architecture. WuDaoKou is a new and complete SoC design."​

​

WuDaoKou - Microarchitectures - Zhaoxin - WikiChip

WuDaoKou is the successor to Zhangjiang, a 28 nm x86 microarchitecture designed by Zhaoxin for mainstream laptops, desktops, and servers.

en.wikichip.org

​

That's the claim, anyway. Yes, they had access to Centaur's IP. ...so, legitimate questions remain.

That's what they say, but I don't believe them.

 

[helper800]

That's what they say, but I don't believe them.

Fair enough.

 

[bit_user]

Fair enough.

Really? How is it fair to claim, on the basis of zero evidence, that they ripped off someone else's design rather than making their own? It would indeed be strange if the former owner of Centaur reused absolutely nothing from them, but perhaps it's just stuff like toolchain and regression tests?

Consider the case of Zen - AMD did a ground-up redesign, not using their own IP. So, there are good reasons why Zhaoxin indeed might have started with a clean sheet. Maybe the CNS designs didn't have the right structures to grow into what Zhaoxin wanted to build? This is potentially knowable though detailed microarchitecture analysis, or if they publicly present such details on their own. There's a real truth here, and paying such heed to naked speculation discounts the degree to which it's discoverable.

Now, it turns out that wikichip is a little bit inconsistent, since they have another article covering the launch of WuDaoKou-based CPUs that claims they were able to confirm some lineage back to Centaur's designs. How much was inherited is still an open question, since the article goes on to claim that:

"WuDaoKou is the first design that is similar to contemporary x86 microprocessors. WuDaoKou finally got rid of the front-side bus (FSB). Previously, the chipset integrated the southbridge and northbridge. In fact, the microprocessor die itself was simply the cores. With WuDaoKou, they moved to a modern SoC design. They also introduced a new uncore which now houses the memory controller as well as all the I/O PHYs and memory and cache arbitration."

Source: https://fuse.wikichip.org/news/733/zhaoxin-launches-their-highest-performance-chinese-x86-chips/
​

Again, that's talking about a CPU they introduced 6 years ago. This article is discussing one that's 2 generations newer. Unfortunately, WikiChip doesn't have an entry for it. It seems the site has stagnated quite a bit, in the past few years.

Anyway, while it does indeed seem like Zhaoxin has benefited from Centaur's IP, I doubt there's much Centaur IP left in these new cores. The main reason I say that is that the above quote shows they already started doing major surgery on the Certaur core, 6 years earlier, and I believe there's no way you can scale up IPC by well more than 2x, without some major refactoring. Again, I point back to Zen, as an example of the scale of changes needed to make such huge strides in IPC.

 

[bit_user]

That's what they say, but I don't believe them.

I respect a healthy skepticism. However, skepticism should motivate us to investigate further, not serve an excuse to be almost reflexively dismissive.

In other words, it should serve as a starting point, not a justification for impassivity.

 

[thestryker]

Anyway, while it does indeed seem like Zhaoxin has benefited from Centaur's IP, I doubt there's much Centaur IP left in these new cores. The main reason I say that is that the above quote shows they already started doing major surgery on the Certaur core, 6 years earlier, and I believe there's no way you can scale up IPC by well more than 2x, without some major refactoring. Again, I point back to Zen, as an example of the scale of changes needed to make such huge strides in IPC.

Unfortunately it's impossible to say with regards to the cores as I don't think there were any products put out with the last Via/Centaur IP thus no die shots. The overall CPU design dramatically changed so I'm inclined to agree that they've largely moved on. There are probably parts that haven't seen much if any change for acceleration and the like. The only thing that makes me question how much improvement they've done is that the performance increases seem tied to core count and clockspeed.

 

[bit_user]

Unfortunately it's impossible to say with regards to the cores as I don't think there were any products put out with the last Via/Centaur IP thus no die shots.

In the article I just linked, they mentioned the predecessor to WuDaoKou had the same floorplan as Centaur's Isaac II. I don't know how they compared them, but it seems like there was some basis for their assertion.

The only thing that makes me question how much improvement they've done is that the performance increases seem tied to core count and clockspeed.

This article is claiming the new 7000-series nearly doubles IPC vs. WuDaoKou's successor.

"Zhaoxin hasn't increased core or thread counts with KX-7000, and the highest configuration remains an eight-core and eight-thread CPU. Assuming Zhaoxin is accurate with its performance claims, the rest of the increased performance presumably stems from architectural improvements, implying a monstrous boost in IPC (instructions per clock)."
​

 

[thestryker]

In the article I just linked, they mentioned the predecessor to WuDaoKou had the same floorplan as Centaur's Isaac II. I don't know how they compared them, but it seems like there was some basis for their assertion.

Yeah I have no idea what they based it on as they don't really say, but I'm guessing it's based on the design diagrams.

This article is claiming the new 7000-series nearly doubles IPC vs. WuDaoKou's successor.

​

"Zhaoxin hasn't increased core or thread counts with KX-7000, and the highest configuration remains an eight-core and eight-thread CPU. Assuming Zhaoxin is accurate with its performance claims, the rest of the increased performance presumably stems from architectural improvements, implying a monstrous boost in IPC (instructions per clock)."​

​

Yeah before this was completely linear with core/clocks, and honestly until someone actually tests the new one I don't buy it anymore than I do when Intel/AMD tout performance. Between the clockspeed increase and addition of a L3 cache there's a lot to boost performance that isn't necessarily tied to a big core improvement.

 

[Albert.Thomas]

Anyway, while it does indeed seem like Zhaoxin has benefited from Centaur's IP, I doubt there's much Centaur IP left in these new cores. The main reason I say that is that the above quote shows they already started doing major surgery on the Certaur core, 6 years earlier, and I believe there's no way you can scale up IPC by well more than 2x, without some major refactoring. Again, I point back to Zen, as an example of the scale of changes needed to make such huge strides in IPC.

If those numbers are accurate, it will put the IPC of those new CPUs almost exactly the same as the unreleased Centaur CNS AI CPUs.

 

[Albert.Thomas]

Unfortunately it's impossible to say with regards to the cores as I don't think there were any products put out with the last Via/Centaur IP thus no die shots.

They actually did release die shots of these CPUs. You can see some of them on the article at Chips N Cheese:

Examining Centaur CHA’s Die and Implementation Goals

In our last article, we examined Centaur’s CNS architecture.

chipsandcheese.com

 

[thestryker]

They actually did release die shots of these CPUs. You can see some of them on the article at Chips N Cheese:

Examining Centaur CHA’s Die and Implementation Goals

In our last article, we examined Centaur’s CNS architecture.

chipsandcheese.com

I totally forgot they got access to one of those after the Intel acquisition of Centaur for some reason.

 

[Albert.Thomas]

I totally forgot they got access to one of those after the Intel acquisition of Centaur for some reason.

I've got one if you'd like me to run some benchmarks 😀

 

[bit_user]

If those numbers are accurate, it will put the IPC of those new CPUs almost exactly the same as the unreleased Centaur CNS AI CPUs.

That's for the wrong manufacturing node, for it to be the same CPU. That doesn't rule out the possibility of a port, however.

We'll have to see what sort of information and analysis comes to light, once the new 7000-series launches.

I totally forgot they got access to one of those after the Intel acquisition of Centaur for some reason.

Who got access to it?

 

[Albert.Thomas]

Who got access to it?

Chips N Cheese, myself, and a few others have access to the pre-release Centaur CHA/CNS AI CPUs

 

[helper800]

Really? How is it fair to claim, on the basis of zero evidence, that they ripped off someone else's design rather than making their own? It would indeed be strange if the former owner of Centaur reused absolutely nothing from them, but perhaps it's just stuff like toolchain and regression tests?

Consider the case of Zen - AMD did a ground-up redesign, not using their own IP. So, there are good reasons why Zhaoxin indeed might have started with a clean sheet. Maybe the CNS designs didn't have the right structures to grow into what Zhaoxin wanted to build? This is potentially knowable though detailed microarchitecture analysis, or if they publicly present such details on their own. There's a real truth here, and paying such heed to naked speculation discounts the degree to which it's discoverable.

Now, it turns out that wikichip is a little bit inconsistent, since they have another article covering the launch of WuDaoKou-based CPUs that claims they were able to confirm some lineage back to Centaur's designs. How much was inherited is still an open question, since the article goes on to claim that:

​

"WuDaoKou is the first design that is similar to contemporary x86 microprocessors. WuDaoKou finally got rid of the front-side bus (FSB). Previously, the chipset integrated the southbridge and northbridge. In fact, the microprocessor die itself was simply the cores. With WuDaoKou, they moved to a modern SoC design. They also introduced a new uncore which now houses the memory controller as well as all the I/O PHYs and memory and cache arbitration."​

​

Source: https://fuse.wikichip.org/news/733/zhaoxin-launches-their-highest-performance-chinese-x86-chips/​

​

Again, that's talking about a CPU they introduced 6 years ago. This article is discussing one that's 2 generations newer. Unfortunately, WikiChip doesn't have an entry for it. It seems the site has stagnated quite a bit, in the past few years.

Anyway, while it does indeed seem like Zhaoxin has benefited from Centaur's IP, I doubt there's much Centaur IP left in these new cores. The main reason I say that is that the above quote shows they already started doing major surgery on the Certaur core, 6 years earlier, and I believe there's no way you can scale up IPC by well more than 2x, without some major refactoring. Again, I point back to Zen, as an example of the scale of changes needed to make such huge strides in IPC.

Because there is no point in arguing with someone who has such an opinion, all that's left to do is respect it and move one. I welcome additional information to form my own opinion though, and for that, cheers!

 

[bit_user]

Because there is no point in arguing with someone who has such an opinion, all that's left to do is respect it and move one.

You don't have to validate it, though. No supporting evidence was provided, and yet we see that a fair bit of information was out there. (seemingly) naked speculation should not be encouraged.

 

[helper800]

You don't have to validate it, though. No supporting evidence was provided, and yet we see that a fair bit of information was out there. (seemingly) naked speculation should not be encouraged.

It may be naive, however, when it comes to CPUs, Albert is among the few I take at their word. I defer to people like him who are much more informed on these matters than I. This is not a subject area I am particularly up to date on.

 

[bit_user]

It may be naive, however, when it comes to CPUs, Albert is among the few I take at their word.

If he had good reasons for thinking so, he should've given us more than his skepticism. Indeed, we see that he had more to offer, on the matter.