It's not use arguing with Jai Russki, he is just living in his own little world. The only downside is that people like him downgrade the discussion level on SDF to fanfiction. Even the Russian government which is notorious for making some propaganda goals and then not meeting them (like head of their state space corporation RKK "Energy" promising to build ) is only going for modest not fully indigenous 28 nm fab by 2030 .
Chinese semiconductor industry
- Thread starter Hendrik_2000
- Start date
- Status
He said nothing about X86. He only talked about how RISC-V was the thing and China was going to do its own fork of the architecture and sell it to the BRICS or something like that. Did you read the same article I did?
Zhaoxin is depending on VIA's X86 license. I think that license expired last year. Wait until either Intel or AMD or both develop new instructions for AI acceleration or whatever and get those patented and Zhaoxin cannot use them.
As is it cannot even run standard Windows 11 because that crap needs TPM 2.0.
As for MCST Elbrus 8C it is a single die processor made at 28nm. So no shit it cannot compete with multiple chiplet AMD EPYC CPUs made at 14nm or better. MCST designed Elbrus 16C made at 16nm with 33% faster clock, twice the memory bandwidth, and twice the cores than 8C. It actually taped out two years ago. Just good luck finding one. And since they were made at TSMC in Nanjing no more production is going to happen. Not with the sanctions. And for those kinds of backend servers running multi-threaded workloads the Russian industry had came up with the Baikal-S processor i.e. a 48 core ARM processor similar to the Phytium also made at TSMC with 16nm. But you can guess what happened to that too.
The license expired last year to my knowledge. Good luck convincing Intel to renew it without suing them. In an US court. And try to convince the court it is a good idea for Intel to license its ISA to a Chinese company. Back when it was VIA in Taiwan and Zhaoxin didn't exist things were different.
I doubt the Elbrus 16C or Baikal-S are behind Zen 1. Even despite being 16nm. Elbrus 16C is capable of 1.5 TFlops. About the same as AMD EPYC 7643 with 7nm process i.e. Zen 3. Elbrus is a really good design for what it was meant to do i.e. peak scientific compute for things like radar signal processing.
I just showed you how a 16nm MCST Elbrus is competitive at scientific compute with a 7nm AMD EPYC. Only architecture with better peak FP than VLIW architecture for same transistor budget would be a vector processor or DSP. And guess what the Elbrus design can also have optional DSP units of ELcore design. It is just that is overpowered for most user applications of Elbrus 16C so it is only in older designs like the Elbrus 2S+. What the Russians end up doing in practice for compute is they use the Elbrus as the control CPU and add Elcore or NeuroMatrix DSP processors around it. At least that is what they do in the Su-57 flight computer. A couple Elbrus-2SM CPUs with several NeuroMatrix 1879VM5Ya DSPs. Which are 64-bit vector processors capable of multiple-precision operations.
The ones that are listed in TOP500. You think the Russians would put something like that on a public list?
Do you think it is a coincidence the Russian machines on TOP500 are owned by Yandex and Sber. i.e. a search engine company and a bank? i.e. civilian companies? Do you seriously think the Russian MIC does not have something better than the civilian sector?
No, they perform way better than a Pentium 3 at the tasks they were designed to do. Also even the best Pentium 3 was at 180nm.
Not that much better. China is still screwed in the commercial sector. ARM would be toast. Even with the Imagination Technologies Chinese owned UK company talked about here. China cannot even manage it and appoint their own CEO. Transfer of IP to outside the UK is highly restricted. China is basically a mostly silent investor in practice.
Last edited:
I am talking about his broader claims about being better positioned to deal with getting cut off from Western chip designs. The talk on RISC-V was just an example.
Zhaoxin is a JV between VIA and Shanghai government, so you are saying that VIA did not renew a license to its subsidiary? I also don't get your second part - VIA's x86 license allows Zhaoxin to produce x86-compatible CPUs, what proprietary stuff of AMD & Intel have to do with that? They are already different.
First of all, who cares about the fairness of comparison, when I am talking about the net absolute of what each industry can offer. You saying that 8C cannot compete with Zen 1 merely proves that China is better suited to withstand chip design denial. Secondly, , its single-thread performance is still far behind i7-2600 (2011 CPUs). The Baikal-S processor you are talking about is also included in one of the tests but I did not find broader tests done for this CPU, so idk how it compares against Kunpeng or EPYC.
What does Intel have to do with VIA's x86 license? There are three companies that hold x86 licenses - Intel, AMD, and VIA (Cyrix). All of them can license them separately. You also did not provide the information about license expiration despite referencing this twice already.
Elbrus single-thread performance is 2x+ behind Intel i7-2600 according to the tests I've provided above while Baikal-S is comparable to AMD EPYC 7351 (Zen 1, 16 cores) according to Baikal itself but there are no independent tests (or I could not find them).
Elbrus 16C's 1.5 TFlops are for float operations according to the MCST itself, while its double-precision (usually used in scientific compute or mathematical simulations for numeric stability) compute is 750 GFlops but I did not find the information about the tests themselves. So if we compared them on HPL, then Elbrus 16C has 560 GFlops according to the test above, which puts it on the level of rather than 7643. Plus, maybe it was designed for scientific computing but eventually, it was positioned as a general use CPU and pushed as import substitution into government agencies & companies.
Nope, you didn't. In order to actually compare performance like that you at least need to specify which tests you were talking about, lol. On HPL (one of the standard ones) Elbrus 16C is on the level of 12-core 3900x, not EPYC 7643. Yeah, theoretically under the ideal compiler (since it is extremely important for VLIW) VLIW is faster than CISC\RISC. However, that's in theory, we are yet to see world-beating VLIW CPUs in practice and the older projects like Itanium were eventually ditched.
I don't get the part about DSPs and other such devices - they are useful for specified tasks (neural network acceleration, fast matrix operations (same shit), etc.) but Elbrus is not positioned as a co-processor. This approach of general use CPU + accelerators is used everywhere, notably in smartphones (for neural net acceleration) - AIPs and DSPs. That does not, however, change anything to the topic we are discussing - the ban on chip designs targets not only MIC but also the consumer and corporate sectors. CPU + bunch of DSPs is not a universal jack-of-all-trades solution.
Oh, not this again - are we getting into the PMC territory with his "Germanic engineering" bs? So you are saying that Russia secretly has some wunderwaffle indigenously designed supercomputer with CPUs that no one knows about despite the fact that they would actually need to fab them outside of Russia? China also stopped putting its supercomputers into TOP500 and similar rankings, we still know about their existence and the fact that China uses indigenously-designed manycore CPUs for them. I am working with facts - China has demonstrated such capacity, Russia has not, and I have zero desire to go into conspiracy bullshit.
The best Elbrus that Russia can fab is Elbrus-2SM (with no DSP block btw) which has 2 cores and a 300 MHz clock, it is worse than the TSMC-fabbed 2S+ (since TSMC's 90 nm is better than Mikron's). Maybe not Pentium 3 (130 nm btw), but around Pentium 4 (90 nm). This does not change the fact that it is useless for the commercial sector and even number crunching.
All the Zen 1 CPUs and Zhaoxins are for the commercial sector, not only for scientific computing, and all of them are safe. How is "China not much better" when you yourself are saying that Elbrus is good at scientific computing while Russian Baikal S was ARM-based (btw, China in that case has Kunpeng 920)? China has both sectors covered with manycores for number crunching and x86 CPUs for general tasks while Russia is trying to make its VLIW-based Elbrus a jack-of-all-trades.
So Ambarella is successful primarily because HiSilicon got banned by the US. Classic Murican behavior when they can't compete.
Even if those instructions are not patented, my guess instructions those are designed to work with AMD and Intel processors as part of they whole instruction set.
I read the Via "offloaded" their centaur team to Intel in a 125 million deal, I don't if this will affect the license to Zhaoxin but I doubt it will affect them.
I think they are getting much better.
With all this GPUs coming from China I ever wonder if they are using their own ISA, Imagination or Verisilicon. Because in most of their websites they don't mention what IP they use, you know like SOC makers do.
Russia spending $60b per year on road construction. Figures on Nuclear are classified but it will be similar numbers considering the export contracts they won. They have railway and railway troops. so hard to pin the budget but infrastructure budget can easily be $20b to $30b a year. so why you think they are not spending similar numbers on semiconductors when in 2030 road map its there strategic field? and Laser is the field they had cooperated longest with East Germany and than Germany. They had spend huge on physics over decades.
Russia simply doesn't have the sales to sustain a competitive semiconductor industry.
My guess is that the Russia accounts for about 2% of global semiconductor demand.
So Russia could subsidise semiconductor fabs and semiconductor equipment, but who is going to buy the output?
A similar situation applies to lasers and other areas.
You do not get it. To have an X86 patent license from Intel you have to fight them tooth and nail for it. And they always put a time limit on it and all sorts of cancellation clauses.
VIA got the license because of legacy licenses from the IPs they held with Centaur and Cyrix. Even then Intel argued it wasn't valid anymore and Intel only lost the case because back then VIA got lucky as Intel was being sued for anti-trust for unlawful behavior against AMD after a pretty big scandal.
The license does not get automatically renewed. It was granted by the FTC as part of anti-trust suit.
"the FTC extended the Via license agreement with Intel for a total of 15 years, it now ends on April 7, 2018. Via now has clear sailing until the patents in question are expired, a huge win for the boys in Austin."
So it wasn't even until last year. The license ended almost 4 years ago. Which means any new patents, any new instructions, Intel developed since then, VIA don't have a license for them.
Man. Did you even read the article you linked to?
"Attention: we are testing an engineering processor with a partially disabled cache, the production sample should have better results."
And like I said Elbrus2K is optimized for technical computing. Look at the MP MFLOPS or HPL GFLOPS numbers. It blows the other processors out of the water in that. The HPL test is basically similar to the test used to make the TOP500 supercomputer index. Of course it won't do well in branch heavy, or, multi-threaded code. The architecture isn't made for that. That is what processors like the Baikal S are for.
It is basically a slower clocked slightly hotter 48 core with like 85% the performance of a similar Kunpeng 920.
More than enough for a bank. Not too bad considering Baikal S was made on 16nm process and Kungpeng 920 was on 7nm.
Who do you think VIA is licensing x86 from? Intel and AMD. Do you think x86 came out from thin air or something? Intel invented the thing. There isn't an x86 license pool, or an x86 Ltd. you can license from. There is the originator of x86, Intel the designers of the 8086 processor, and then there are all the patents made on x86 after that to this day. Hundreds of them. Intel has patents, AMD has patents, even VIA has patents with things like Padlock, or AES encryption, but no one else uses them.
It is BS, because he is comparing a non optimized BLAS, no one in the field will use that. They will use an hand tuned manually written assembler coded optimized version of it. Which MCST likely charges money for. As does Intel for theirs.
Yeah that happens too. It is not like MCST pushes people to use Elbrus2K for everything. And like I said there are other options. At least now there are. Or were. Even MCST clearly knows Elbrus2K isn't for everything which is why they have the SPARC processors like the R2000.
You know, there are other VLIW architectures than Itanium. Philips used to have the TriMedia processor for example. It was pretty popular back in the day to encode and decode video at a time other CPUs couldn't do it at any appreciable speed.
TriMedia would be the kind of thing you would find inside a DVD or Bluray player.
The MIC uses CPUs for technical compute. They do not run SQL databases and Java servers on weapons systems.
We are talking about real-time or as close to it as possible complex computations of trajectories, signal processing, neural networks, and crap like that.
Last edited:
It is like this. There are catalogs of chips. Yet it seems like there is nothing on commercial systems or system offers which uses them. The government funded the development. The architecture was originally designed for the MIC. So, what do you think. Do you think they just design CPUs, and compilers, and make a test batch with one wafer. And call it quits? Because you know what. The weapon systems only use the 90nm processors made in Russia. They don't use this 28nm or 16nm stuff. So where is it being used? Wafers are typically ordered in thousands minimum at a foundry. Each wafer can fit at least a dozen CPU dies in it. Just use your brain a bit.
It blows the PowerPC processor used in the US F-35 flight computer out of the water. That is enough for the MIC.
The Su-57 flight computer has boards with two Elbrus-2SM chips and like 4 DSPs chips in it.
And it is quite useful for FP compute. The internet cloud sector and web services were never the target of this.
So you are telling me the Russian MIC should be buying Zhaoxins fabbed at TSMC? Yeah that would work. Not.
Like I said. Russia has Baikal S for crap like that. And there are people working on RISC-V in Russia too.
You seem to be obsessed with x86 compatibility for whatever reason. I mean even Apple doesn't care about it.
The MIC write their own software for the weapon systems. They don't use other people's.
Is there any doubt the real reason Huawei got banned was competition to Apple and Qualcomm? It was only when Huawei got competitive in the smartphone market that they came up with the BS they had to kill it because of security issues in every single system they made.
Typically the instructions are not patented per se. They patent the way to implement those instructions in hardware. The way they did it, and all the other ways to implement them in hardware they could think of. Just to f u over.
It used to be that the Chinese team at Zhaoxin picked up the Centaur design and tweaked it a bit. Theses latest processors supposedly are their own full up design. But I have also heard they reused some components and shuffled things around and added some bits. Who knows. We need more info.
IIRC one of the Chinese GPU vendors licensed the PowerVR GPU architecture from Imagination Technologies and slapped it in a chip. The other two supposedly made their own designs.
Last edited:
- Status