Amid the push for technology independence, Chinese companies are pushing out more products to satisfy the need for the rapidly soaring demand for domestic data processing silicon. Today, we have information that Chinese Loongson has launched a 3D5000 CPU with as many as 32 cores. Utilizing chiplet technology, the 3D5000 represents a combination of two 16-core 3C5000 processors based on LA464 cores, based on LoongArch ISA that follows the combination of RISC and MIPS ISA design principles. The new chip features 64 MB of L3 cache, supports eight-channel DDR4-3200 ECC memory achieving 50 GB/s, and has five HyperTransport (HT) 3.0 interfaces. The TDP configuration of the chip is officially 300 Watts; however, normal operation is usually at around 150 Watts, with LA464 cores running at 2 GHz.
China’s rapid improvement in microprocessors isn’t really all that interesting for us in other parts of the world, because chips from companies like Loongson don’t really make their way over here. What is interesting about this, however, is the implications this continued trend will have for the geopolitical state of the world. A China not dependent on Taiwan’s TSMC for its chips is a China that can more freely invade Taiwan.
We’re even less likely to get ahold of one now…
https://www.tomshardware.com/news/us-govt-blacklists-loongson-and-inspur
Who manufactures these CPUs?
It seems to be done on a 12nm process, so SMIC is likely fabbing them.
It’d be worth noting that the interest in TSMC is for high tech fabrication, not chip design. These chips are being manufactured on *extremely* out of date processes based on the quoted power draw, since a comparable phone ARM chip is in the 5W range, compared to the 300 in the article. They have a long way to go to catch up, it’s not just about core count.
12nm is not “extremely out of date.”
It’s not the latest node, but for the intended application they seem to be good enough. Since they seem to be reaching Zen 3 performance, which is not too bad for a fully homegrown product.
I see nowhere they claim zen 3 performance..
they claim spec2006 of 425… but dont’ mention which test.
The only test that has similar numbers for a Zen system are passed up pretty far by even AMD’s Naples Zen 1 7601 at half the power… An Eypc 7601 gets over 1070 in CINT2006 and 871 in CFP and that is a 32core CPU in single socket config.
Zen 3 and 4 would be way past this…
https://wccftech.com/chinese-built-loongson-3a6000-cpus-achieve-68-percent-faster-single-core-performance-will-rival-amd-zen-3-intel-tiger-lake/
You do realize that nm numbers have been pure marketing for a long time, right? Based on these power draw numbers, this “12nm” process is nowhere near what TSMC’s equivalent was.
Every product name has always been a result of marketing. I have no clue why some of you genuinely think this is a new development or that semiconductor processes were an exception.
FWIW node numbers have traditionally referred to the smallest discrete resolution capabilities for the lithographic system. These being no different. And I still don’t understand why people not involved in the design and fabrication cycle would have any concern with them.
FYI I work in this industry, and I literally interact with process details on a daily basis. We have no data points to compare against the equivalent TSMC process from just the expected power draw from a random product. We don’t even know exactly which process has been used.
javiercero1,
It’s because everyone wants to know where the Chinese really stand on fab capability and process technology is a significant indicator of this.
Yes, everyone’s on the same page with wanting to have more info & specs. Additionally it’s important that performance and power consumption be independently verified since manufacturers exaggerate. Alas, until more specs and empirical data arrive, most people’s questions aren’t going to be answered.
Everyone? You overestimate severely the amount of people who give a hoot about China’s semiconductor technology, or any other country’s for that matter.
javiercero1,
Yes, obviously I’m referring to everyone here, including both of us, who would like to see more details.
Samsung taped out 14nm in 2014 and 10nm in I believe 2016. So let’s split the difference and say a 12nm node size is circa-2015 technology. So it was a leading edge node 8 years ago. I’d say that’s pretty out of date.
Once they hit 128 cores then they should be in a strong enough position to declare war against Taiwan I think. Even if these chips are manufactured with a 200nm process from the 90s.
Noticeably absent: The process (nanometers) it’s built on.
12nm
So, around where TSMC was in 2018 when they were making Nvidia’s Turing cards. That’s better than I expected.
SMIC has already done risk runs for sub 10nm processes. So I’d say they seem to be around 5 years behind. Which is impressive if you realize they don’t have access to IMEC and ASML
It’s not as impressive when you realize they’re using DUV technology and they have literally hit the limitation (around 7nm). Even with the constant theft of IP it’ll take them decades to re-invent EUV litography, which is how long it took the rest of the collective world.
So with Loongson they have MIPS/ARM and with Zhaoxin they have X86 so it will be interesting to see what they come up with.
Well, with Mediatek and Huawei they have ARM as well 😉
They also have RISC-V parts on several bring up stages withing a few orgs in China.
There are a couple of homegrown GPU/TPUs that are hitting their internal market as well.
And SMIC will soon have capacity for sub 10nm process that are leaving the risk runs.
Their main achiles heel is that they lack a full EDA flow, although they have a few startups doing some good tools.
It is remarkable how quickly China has been able to become somewhat self sufficient in terms of computing.
Nope, when it comes to the high performance-per-watt stuff, they are dependent on TSMC and by extension the United States. It’s the same reason Russia has to “side-import” chips for their missiles, because it’s impossible from them to make missiles using whatever lithography they and their buddies in China have available.
And TSMC is, of course, dependent on ASML, which is in Netherlands.
But can China make a desktop or laptop, even with vastly inferior performance-per-watt? Yes, they can. I guess that’s what their goal is.
Everybody is dependent on TSMC for the best fab node currently, and that applies to the US as well.
But in order to have a homegrown computer infrastructure, that can is basically done bootstrapping, it is pretty impressive what the Chinese have been able to achieve in less than 2 decades. SMIC is also completing sub 10nm capabilities, so the gap is closing. when it comes to their own mobile processors.
> it is pretty impressive what the Chinese have been able to achieve in less than 2 decades
Not when you realize all the technology was stolen.
This is exciting news! I was one of the first few that ordered the first Gen Lemote Loongson mini PCs. Ran OpenBSD 4.3 on it. Too bad there’s a trade embargo on Computers from China.
According detailed write up (https://chipsandcheese.com/2023/04/09/loongsons-3a5000-chinas-best-shot/) this is roughly performance equivalent of Sandy Bridge while in some aspects it comes out worse than equivalent chips that were designed on equivalent node and for the same operating frequencies, so we’re talking about ~ 10y gap.
The main question is how is the destkop/server cpu lead translating into military advantage. Latest gen chips are not needed in rockets to home targets. What counts here are radar, rocket technologies and algorithms.
The node difference will have more impact on AI research where each terraflop counts for training performance.
That is indeed a nice in depth comparison… with none of the nonsense. I have no idea where people were thinking this was actually as fast as Zen 1 when it lags behind it considerable. I’m surprised they didn’t bother to upgrade to perception branch prediction…. to me that would imply there isn’t as much R&D going on with this design and its just being rote spec upgraded with little improvement.
I mean china didn’t even bother to steal the technology to make this chip better… its just lagging behind.
They are reaching IPC parity with Zen3. Which is impressive considering.
What is a “perception branch predictor?” Did you mean a perceptron?
Auto correct and no… they aren’t reaching IPC parity at all. They are about 2-2.5x behind as far as I can tell.
And yes I meant perceptron, autocorrect being overzealous there.
You can’t tell much, I am afraid.
Perceptron-based predictors are ancient BTW.
Just to be clear, milspec electronics tend to be several generations behind the consumer stuff in the west as well.
E.g. the F-22 until recently used i960s sourced from the 90s.
What? I’m writing to congress. They need to be on the latest gen and latest nodes, at all costs, at all times! It’s a travesty the last red, white, and blue fighter can’t run the latest AAA games titles. The troops must be so humiliated no wonder we haven’t been able to take credit for winning a war since 1945.
I know that you wrote in Jest. But usually anything milspec and aerospace is way behind because it has more strict verification/validation and sourcing requirements. This also applies to stuff like medical equipment, stuff like recent MRI controllers, etc still run on 486s.
Which means that from a technological perspective the Chinese military has reached some of their goals of having a fully home grown supply chain. 12nm nodes is good enough for these types of applications, and they have that flow mostly detached from Western suppliers. So sanctions are going to have less strategic effect there, specially when it comes to supply Russia.
Yeah. 🙂 This is fine for what it is, and hardened, industrial tech is a totally different realm then regular tech.
The second paragraph makes a good point that everyone may be missing.
javiercero1,
I think this depends on specific “military applications”. As for conventional CPUs that go into a drone, or bomber, or tank, or submarine, etc. Then sure I agree with you “12nm nodes is good enough for these types of applications”. However another significant type of military application clearly lends itself to even smaller node sizes: code breaking. A better process significantly shapes viability and reduces operating costs. Military AI is another application that would benefit from smaller node sizes. China today may still be dependent on illicitly sourced western GPU/tensor acceleration hardware.
I think many of the sanctions are less strategically effective than advertised for various reasons. We’ve got many gray & black markets that the US doesn’t have a handle on even at their own borders much less throughout the world.
I think brute force code breaking hasn’t been practical for a while, with current crypto systems. Thus the insistence of backdoors.
China has already managed to have a homegrow system in the top-10 supercomputers.
And they’re almost reaching tensor parity with their Biren GPUs.
They’re still not in the lead in terms of power/performance. But they’re getting there in terms of performance. Which again, it is impressive given where they were a couple of decades ago.
javiercero1,
It depends. Crypto analysts are regularly finding mathematical weaknesses in algorithms, but the more immediate danger is strong crypto protecting data with too little entropy in the first place such that the search space is significantly compromised. This is notoriously the case for password hashing and it renders brute force attacks quite viable. Salting increases the scaling costs but hardware improvements are making it more viable to brute force single passwords with no rainbow tables. Cryptographic hash based authentication is prevalent in technologies everywhere from active directory, laptops, websites, email, wifi, etc. If they intercept a device or glean a hash through an internet protocol, there’s a non-negligible risk that a nation state can brute force it. Thanks to birthday attacks, biometric hashes are also vulnerable to brute forcing.
Another thing to consider is that China may consider enemies of the state quite differently than you or I would. A civil protest in hong kong might be something that china would want to squash using NSA-like resources for reconnaissance. A pesky reporter could be target. IMHO it’s ridiculous to consider these enemies of the state, but arguably still on brand for them.
Yeah, though I’m not sure that would be true if the lists included the super computers of government agencies around the world. If they weren’t so secret & obscure, these would likely bump everyone else down the list, but all the details are classified so we don’t get the benefit of knowing 🙁
Do you have a source comparing Chinese power & performance specs to other commercially available CPUs?
The link you gave earlier talking about loongson 3a6000 is interesting, but every article I find on the CPU seems to be written from the same self-promotional press release from last year that talk about the processor specs strictly in future tense (“is expected to”, “will be”, “should be”, etc). I haven’t seen a single report cover the actual silicon performance. I’m genuinely curious where chinese capabilities stand, but personally I need to see something more concrete.
Over the years I’ve learned that it’s not wise to let manufactures score themselves because they often abuse the privilege, haha.