News Chinese chipmaker claims new Loongson 3B6600 CPU could hit 13th-Gen Intel performance

My guess is that they're basing these statements on SPEC2006 or other obsolete benchmarks. That's what I recall seeing, last time these sorts of claims came to light.

I do wonder what sort of lithography they're using, as that should have a lot to do with the amount of IPC they can achieve.

FWIW, I tried following the source link, but there's no more information on there than what the article covered.
 
If anyone is curious how the predecessor compares to modern x86 cores, here's a benchmark of its single-thread performance:

https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93750689-49f2-406a-80c8-cc6b51a33dc9_1226x713.png


Source: https://chipsandcheese.com/p/running-spec-cpu2017-on-chinese-cpus

Here's a detailed analysis of the 3A6000:

The author is somewhat complimentary:

"Engineers at Loongson have a lot to be proud of. Creating a branch predictor on par with Zen 2’s isn’t easy, and that achievement is made all the more impressive considering where Loongson was with 3A5000. Similarly, SMT is very difficult to get right. Loongson managed to do both while dramatically scaling up the 3A5000’s out-of-order engine and fixing its DDR4 controller. The resulting massive performance gain puts LA664 on par with Zen 1, at least when each core only has a single hardware thread active."
...
"Zen 1 level single threaded performance is commendable. But we have to remember Zen 1 gained market share against Intel because it brought affordable 6 and 8 core parts into consumer platforms, not because it could win core for core against Skylake. The 3A6000 is just a quad core part, and thus lacks Zen 1’s biggest strength."
 
My guess is that they're basing these statements on SPEC2006 or other obsolete benchmarks. That's what I recall seeing, last time these sorts of claims came to light.

I do wonder what sort of lithography they're using, as that should have a lot to do with the amount of IPC they can achieve.

FWIW, I tried following the source link, but there's no more information on there than what the article covered.
Do you remember what the predecessor's lithography was? I definitely agree that it could have a lot to do with a better process.
 
All very interesting.

It's good (IMO) to see another independent CPU maker get better at their craft. They are not yet as big as the current giants in the market for sure, still they keep getting better and are making progress fast. It'll be fascinating to see if/when they catch up to AMD/Intel (or even ARM).

Of course the fear of some kind of Chinese Government mandated back door or security hole would likely keep me from ever using one for anything beyond a calculator and maybe not that.
 
  • Like
Reactions: Findecanor
When you look at the whole picture, not just desktop chips, Intel 13th gen isn't impressive in the least.

The top dog is the AMD EPYC 9965.
If the Longsoon starts beating a Xeon Max 9480, then sure, let's have a look.
 
I do wonder what sort of lithography they're using, as that should have a lot to do with the amount of IPC they can achieve.
Loongson is a public company (SHA:688047), so they have to hold shareholder Q&A sessions, like this one:
https://www.bilibili.com/video/BV1NNDoYeEK6

in this Q&A, the company owner claimed that they used the "mature node" which "comparable to 7nm x86". Loongson's principle is to use the node that has a domestic backup, so they won't be sanctioned to death as a chip design company.

so my guess would be between 10~12nm. their 3A6000 was 12nm.
 
  • Like
Reactions: bit_user
The biggest hurdle of Loongsong is likely the software ecosystem which I don't see how they can overcome easily. Developers can port over Firefox or chrome but will it have the most optimized code and renderers/codecs? They need that at the minimal to be the "chromeos" of China. To run more mainstream software , it will also have to do binary conversion and emulation like wine.
 
I do wonder what sort of lithography they're using, as that should have a lot to do with the amount of IPC they can achieve.
Almost assuredly is a variation of 193nn immersion lithography (duvi) and probably using self alignmened quadruple patterning. That's what the other fabs use to get "7nm" nodes without euv