Japanese CPU threatens to make Nvidia, Intel and AMD obsolete in HPC market

And even more so from an old traditional Japanese enterprise like Fujitsu.

At most, they designed the entire CPU from the ground up. I'm guessing it's somewhere in between; this CPU provides equivalent extensions to what Fujitsu gave in their SPARC series. If other parts of ARM's core IP were suitable for other pieces they probably used that.

> Who manufactured it?

Based on Wikipedia, probably TSMC.

> After the us tries to cut off Huaweis chip supplies, can Huawei come up with something similar?

(Short answer at the bottom, I went on quite the tangent here :)

Well let's talk about what that would mean right now.

HiSilicon is a fabless Semiconductor design company fully owned by Huawei. If you look at the comparison of ARMv8-A cores out there [0], thankfully both HiSilicon and Fujitsu's offerings are grouped together at the bottom (at least for me as I write this.)

That said, what benchmarks I found indicated that HiSilicon's CPU doesn't seem to compete, especially with wider workloads (Fujitsu can do 512-bit vectors, I think HiSilicon is limited to 128 on these.)

I'm assuming that their setups are being pieced together more thoughtfully; one should note HS also makes AI Processors that are tailored separately for inference and training.

As far as making a more proper 'server' GP core, if they're not playing dirty there's some catch-up to be done, but there's nothing stopping the State from stepping in and throwing more resources at the problem. While having multiple design teams working on the same parts of a CPU probably won't get an implementation done any faster, having multiple designs in parallel when you're trying to catch up to the latest tech may lead to better insight and options down the road; case in point would be Netburst vs Evolved P6 arch.

Really, China's bigger problem at the moment is Fab technology; They have HSMC which can do 7nm, but I don't know at what kind of volume, and whether it's a bulk process or something with all the fancy tech you'd need to make a server cpu at a fast clock rate.

There's a few things they're probably lacking in that area; just getting started into process tech research (in the vein of SOI, copper interconnects, etc.) but that will probably ramp up quickly.

Optics are another huge factor; I'm starting to realize that the increase in higher-quality Camera glass from China in the last few years is no coincidence. Let's look at the main Lithography EQ manufacturers out there:

- Canon - Nikon - Ultratech - ASML

You can see that two of ASMLs biggest competitors also happen to be the two biggest/best mass producers of Glass out there. Notice Samsung isn't on the list; AFAIK they still get their equipment from ASML, although my understanding is it's some level of 'partnership.' I'm guessing they trade glass, since Samsung is pretty good at that.

To wit: Samsung recently got caught with their hand in ASML's source code cookie jar. I was immediately reminded of a post on HN [1] where ASML's testing scheme is described; If someone has a spare machine and a copy of the source code, they can start playing with it, possibly introducing their own fixes to optimize yields, either for their use case or for more general cases that for reasons described in aforementioned post aren't incorporated into ASML's codebase.

Anywho, went on a tangent there, but it shows how competitive the industry is even without political pressures at play. But it also shows that it's -possible- that with the right pieces China could make incremental improvements to what they already have while they work on the next big thing.

Short Answer:

I think China wants to be caught up sometime between 2023 and 2025 for fabs, but everyone else is trying to push it back to 2025-2030.

[0] - https://en.wikipedia.org/wiki/Comparison_of_ARMv8-A_cores

[1] - https://news.ycombinator.com/item?id=18463181