Loongson's 3A6000 Brings Simultaneous Multi-Threading To LoongArch

[phoronix]

Phoronix: Loongson's 3A6000 Brings Simultaneous Multi-Threading To LoongArch

Since last month Loongson engineers have begun posting Linux patches enabling their upcoming 3A6000 series LoongArch processors under Linux. Yesterday they posted new patches and revealed that Loongson 3A6000 processors support Simultaneous Multi-Threading (SMT)...

https://www.phoronix.com/news/Loongson-3A6000-SMT

[stormcrow]

Wonder how much of 3A6000 and its surrounding support tech is stolen technology at this point.

https://www.reuters.com/technology/south-korea-indicts-ex-samsung-elec-executive-alleged-data-leak-china-2023-06-12/

https://www.bbc.com/news/world-asia-china-64206950

https://apnews.com/article/barracuda-mandiant-cybersecurity-china-hackers-a52d1595c9108d2c58df11e38756600d

etc...

[ryao]

Wonder how much of 3A6000 and its surrounding support tech is stolen technology at this point.

https://www.reuters.com/technology/south-korea-indicts-ex-samsung-elec-executive-alleged-data-leak-china-2023-06-12/

https://www.bbc.com/news/world-asia-china-64206950

https://apnews.com/article/barracuda-mandiant-cybersecurity-china-hackers-a52d1595c9108d2c58df11e38756600d

etc...

It will probably be fabricated using a process made by pouched TSMC employees, but I would be surprised if the processor itself had “stolen technology”. That being said, it is often impossible to make anything without touching some sort of patent, so I would not be surprised if there were unintentional infringements. Google, Apple, Intel and others unintentionally infringe on other people’s IP all the time:

https://hbr.org/2022/08/big-tech-has-a-patent-violation-problem

Loongson in particular tends to cite normalized per MHz numbers to make their processors look competitive, but reality is different. First, performance per clock decreases as clock speeds increase because of the von Neumann bottleneck, so you can measure impressively high per MHz numbers by measuring at a low clock speed, but they are not comparable to normalized measurements of competitive processors operating at much higher clock speeds. Second, their designs do not reach high clock speeds. Additionally, this review is highly critical of the previous generation:

https://chipsandcheese.com/2023/04/09/loongsons-3a5000-chinas-best-shot/

I do not think that loongson reached where it is by stealing (they properly licensed MIPS), but I do think that the capabilities of their hardware are exaggerated.

[stormcrow]

It will probably be fabricated using a process made by pouched TSMC employees, but I would be surprised if the processor itself had “stolen technology”. That being said, it is often impossible to make anything without touching some sort of patent, so I would not be surprised if there were unintentional infringements. Google, Apple, Intel and others unintentionally infringe on other people’s IP all the time:

https://hbr.org/2022/08/big-tech-has-a-patent-violation-problem

Loongson in particular tends to cite normalized per MHz numbers to make their processors look competitive, but reality is different. First, performance per clock decreases as clock speeds increase because of the von Neumann bottleneck, so you can measure impressively high per MHz numbers by measuring at a low clock speed, but they are not comparable to normalized measurements of competitive processors operating at much higher clock speeds. Second, their designs do not reach high clock speeds. Additionally, this review is highly critical of the previous generation:

https://chipsandcheese.com/2023/04/09/loongsons-3a5000-chinas-best-shot/

I do not think that loongson reached where it is by stealing (they properly licensed MIPS), but I do think that the capabilities of their hardware are exaggerated.

No, I think it's quite likely at least some of the technology in those chips is poached as MIPS processor technology is pretty ancient even at the point they stopped pushing it and their corporate parent opened up MIPS architecture for royalty free use. China doesn't need a license to produce MIPS based architecture CPUs because Wave Computing deprecated and opened up the architecture in 2018. But that doesn't mean they have the modern fabrication facilities to pursue the final product. You can't design a practical CPU without taking into account what it looks like on the wafer, that's what they're stealing from high tech companies, the plans for fabrication tools and facilities, and that's why I wonder how much of that finished product was actually designed by Samsung, Intel, AMD, TSMC, etc.

[tuxd3v]

wow...loongson starts to flex its muscles..I think they understand now, that...its a make or break situation..
We just need to remember that riscv designs from china are out there and progressing at break neck speeds..

I agree with Michael, we should be able to get performance metrics from their SMT 2 implementation, and other metrics to see the progress they are making in this arch.

About the claims many are doing above, of industrial espionage..well almost everybody is doing it, the US sits above the cake...is it wrong?
Hell yes it is, but in the broken world we live where rules of respect are no longer valid...will the rest of the world play by the those rules, when some steal everything they can?
Hell no..they wont.

My personal opinion is that at some point in time..what we are experiencing.. is unfortunately a great receipt for disaster.

[ryao]

No, I think it's quite likely at least some of the technology in those chips is poached as MIPS processor technology is pretty ancient even at the point they stopped pushing it and their corporate parent opened up MIPS architecture for royalty free use. China doesn't need a license to produce MIPS based architecture CPUs because Wave Computing deprecated and opened up the architecture in 2018. But that doesn't mean they have the modern fabrication facilities to pursue the final product. You can't design a practical CPU without taking into account what it looks like on the wafer, that's what they're stealing from high tech companies, the plans for fabrication tools and facilities, and that's why I wonder how much of that finished product was actually designed by Samsung, Intel, AMD, TSMC, etc.

China is a big country. Just because some people did industrial espionage does not mean that Loongson did. The 3A6000 is the result of 20 years of R&D. It highly likely that they reached this point on their own efforts.

They are also a MIPS licensee that predates 2018 by many years and they are a fabless company, so talk of stolen fabrication tools and facilities does not apply to them.

Modern CPUs are designed in verilog and verilog does not care about how it is laid out on a wafer. Such details are determined when synthesizing the design. For example, none of the CPU cores here were designed with any knowledge of how they would be laid out on a wafer and they are “practical” CPUs:

https://github.com/enjoy-digital/litex

That being said, a number of your remarks suggest to me that you have no clue:

* Making the mistake that some Chinese firms doing industrial espionage means every Chinese firm is doing industrial espionage.

* Thinking that a company that has been doing CPU R&D for 20 years (now 21 years) cannot build what they built. There are startups that have achieved better results in far less time. The PA Semi guys that designed Apple silicon would be an example.

* Ignoring their MIPS license because of a purported event in 2018.

* Accusing a fabless company of stealing fabrication tools and facilities.

* Talking about things that simply do not exist in the design stage as part of CPU design.

[Estranged1906]

So let's see the current state of CPU architectures:

Actually used

• x86_64 a.k.a. amd64
• aarch64 (ARM)
• ppc64le (POWER)
• s390x (IBM Z)

Coming soon (TM)

• RISC-V
• LoongArch (MIPS-based)

Probably dead but maybe not

• Elbrus 2000 (e2k) - Russian-developed, is this still a thing since the war and sanctions started?
• SPARC - Fujitsu announced end of production soon but it's "open and royalty-free" and there's been a few open source implementations so theoretically someone can go and make a SPARC CPU right?

[tuxd3v]

So let's see the current state of CPU architectures:

Actually used

• x86_64 a.k.a. amd64
• aarch64 (ARM)
• ppc64le (POWER)
• s390x (IBM Z)

Coming soon (TM)

• RISC-V
• LoongArch (MIPS-based)

Probably dead but maybe not

• Elbrus 2000 (e2k) - Russian-developed, is this still a thing since the war and sanctions started?
• SPARC - Fujitsu announced end of production soon but it's "open and royalty-free" and there's been a few open source implementations so theoretically someone can go and make a SPARC CPU right?

Probably longarch is more used than any of IBM products..we just need to remember that China has a population around 2x of the EU and the US combined..
So you can put it in the list of already used archs..
Elbrus is also a already used arch...it always was, but now its in datacenters, and public education institutions.
For what I see,
Russia is developing its own lithography machines..it will take a lot of time, 10 or more years..in the meantime, they can build it in friendly countries(probably at very high cost..), but I don't think China is interested in producing it...after all..China has longarch to sell right?

Sparc is indeed a very good arch,with very powerful security mechanisms, but it takes a lot of money to develop successful products out of it for the datacenter..
And no..it wont be Oracle to do it...because otherwise they will loose the x86 market..it needs to be someone else..
Japan has designs based on it, Russia also, EU also(at least sparc v8-leon processor)..and the US ofcourse also.

So sparc still continues to be used, in aviation, in space, and several types of appliances, however, I don't see it now in datacenters, since no one is interested in develop it further..its a shame.
For what I understand it will be a arch delegated for critical things, like aviation/space,etc.

Riscv...its already in a lot of places too, EU already has it in space, and it will substitute the sparc v8 design EU used for aviation/space, Russia/China are already using it massively in microcontrollers, and trying to advance it to the linux world, however, I China is a bigger powerhouse with other ambitions, and of course, is developing it as a higher pace than any other country, we just need to remember that China didn't had other options.
In the near future we will have 2 laptops based on it..I just think their V and B extensions ...are still draft based and not ratified versions, maybe I am wrong..

You Forgot about Apple own arch, using ARM instruction set.

[aonix]

Loongson used to do multi-core even more than intel, could it be that intel sent spies to steal things from Loongson?

[AndyChow]

Curious if it's really 8 core. The 3A6000 was supposed to be a 16-core chip, not 16-thread. The 3B6000 was supposed to be the 8-core variant with big.LITTLE 4 efficient cores and 4 performance cores.

From the old (and now obsolete?) roadmap:

Loongson is targetting the first 16-core 3C6000 chips in early 2023 followed by 32-core variants in mid-2023 while the next generation will follow up a few months later in 2024 with the 7000 lineups, offering up to 64 cores.

(source)