Loongson Technology develops its own CPU instruction set architecture

Fortunately, RISC-V is really strong in China.

This decision by one CPU vendor is unlikely to change this.

> possible national security applications

I wonder how much of a benefit that would be.

To have an entirely separate instruction set and all the hardware / software that builds upon it that is only really used in your country _sounds_ like it would be a good way to frustrate espionage / hacking to a layman since adversaries lack the expertise.

I am sure adversarial agencies would appoint a person or a small team to become experts in this specific technology though so maybe not so much although the costs would be more than for a "common" tech.

Interesting theoretical situation to play with though.

Per this comment

https://www.cnx-software.com/2021/04/17/loongson-loongarch-c...

It seems risc-v relies on a patent pool dominated by us companies. So the us can arbitrarily extrajudicially stomp on chinese chip manufacturer as they did with Huawei

RISC-V is being widely adopted in China, just not by this one company. I'm sure they have their reasons but the decision seems to make little sense for Loongson. They'll have to port the toolchains and kernels (which was a multi-year effort for RISC-V). If they needed extended instructions for specific features, then RISC-V is ideal for that since it's built around the concept of extending the ISA.

RISC-V was OK before the 1.1 spec. Version 1.1 is self-sabotaged by the RISC-V team, in such a way that you can't create a cycle-accurate software emulator running at reasonable speed. The dirty deed is in changes in the instruction encodings, multiple bit shuffling instructions in the emulator are now required to decode some 1.1 insns.

To me, this indicates that the RISC-V team has bad intentions, and I dropped RISC-V for this reason.

Link? Very interested in learning the details.

Yes. If you design the ISA carefully, you can create a microarchitecture for it that is much more performant and simultaneously more energy efficient. This is one of the reasons that Intel struggles to compete with the Apple M1. The latter does not have baggage of backward-compatibility that Intel is forced to support. For specific use cases, you can get even more performance by tailoring the ISA to the application. E.g. many neural processors fall into this category, their SIMD instruction sets are carefully chosen to be most effective for the sort of calculations you will need to do when you build a neural network.

Loongson will probably stay fed with their military contracts. But just don't expect anything good to come out of the commercial front: they are known to never provide any useful documentation, with the Chinese Loongson user/hobbyist community itself relying on leaks.

With inexistent documentation comes an awful open-source toolchain. Their official community "Loongnix" cgit is a collection of extremely outdated, long-winded forks of gcc and the kernel. They do provide patches to GCC and Binutils, but only barely enough for the errata (and the old MMX copycat). LLVM is overlooked here, so good luck with that ll/sc lockup when you try Rust. [The "grassroots" @loongson-community on GitHub does try to figure out some more instructions.]

And that brings me to the topic of their fancy ISA extensions. Accelerated X86/ARM emulation, 256-bit SIMD (with a mnemonic of "VBCMULADDPD"), everything! But to hobbists who bought their commercial samples, all that silicon area is just expensive padding with no way to use them.

If you can tolerate Google Translate, I recommend reading https://zh.wikipedia.org/wiki/LoongISA and some of the refs. Especially the comment article by "windows1089" on Zhihu.

Approximately 15 years ago svcd was wildly successful throughout Asia and SE Asia, in the 100% piracy market, when lots of people could afford a cd-r writer but dvd-r writers were still cost prohibitive.

SVCD uses MPEG2 too so in the end it's close to a DVD.