Chinese Loongson chips coming in 2023, on par with 2020 x86 kit China is set to get its hands on homegrown processors next year that purportedly rival the performance of AMD and Intel chips released over the past two years Chinese semiconductor company Loongson recently announced that its next-generation Godson CPU, the 3A6000, will sample with customers in the first half of 2023, …
On other sites, the "on par with an AMD Ryzen 5000" is based on IPC - the expected clock rates are in the 2GHz-2.5GHz rate, so likely half the performance assuming this is per core.
For previous generations of Loongson, performance tended to match x86 in a subset of tests and general performance tended to be lower than initially advertised
The real question is if this can be produced at the supposed 7nm process node in volume - previous versions at "new" SMIC process nodes yielded very poorly so 3 years may be closer to 6-8 years. If they have a working "7nm equivalent" process node that has acceptable yields and the necessary equipment to produce these in volume, then it is indeed impressive even if the performance is less than stated as that would likely provide a competitive high performance Chinese CPU to allow an alternative to western equivalents
From a translation of the linked website "According to Loongson’s previous information, 3A6000 will not continue to improve the process, and will still use the existing 12nm process, but will greatly improve the architecture design. The architecture will be upgraded from the current GS464V to LA664, so the single-core performance will be greatly improved". The current 3A5000 were made at TSMC fab using their N16/N12 process, so the question is where will the 3A6000 be made China at SMIC ?
It also matters how much power the devices consume and consequently how much heat they produce in achieving this performance. This information is not in the article but if these processors are efficiently relative to similar performing x86 then I’m sure a song and dance would have been made of it.
Where would CCP add this feature ?
In the instruction set ?
Some super secret extra CPU core dedicated to this feature ?
The binary firmware which is not currently encrypted, so it could in theory be backward engineered and fully audited ? (The beta releases even contain the debug symbols!).
In the binary firmware blobs used to provide access to additional hardware ?
In the Operating System (Just like Microsoft:telemetry, Google:everything and Apple:everything) ?
The ideal access for x64 is via the Intel ME and AMD PSP both of which run signed and encrypted code that can never be audited. All it would take is one FISA court order for the feature to be enabled/added.
I'm not saying that China wouldn't, just that the simplest tap point is not the hardware but the software. And since this is designed mainly for the internal market, the OS is the cheapest and easiest place to add and continually update any feature like that.
If the performance is even close to that being trailed then this is a big win for the Chinese
equivalence to 2020 era AMD64 CPUs and on a MUCH larger die process....
If that really is true then it can be assumed that parity in performance has been achieved already
One minor thing is that the Loongson 3A6000 chips will be manufactured with a 12 nm process. Will that be the exact same 12 nm process as the Loongson 3A5000 (at TMSC), or a later iteration of the 12nm TMSC process.
The simulations of the 3A6000 are being compared to chips made with the 10 nm Intel SuperFin process (100.76 million transistors/square mm) and the 7FF TSMC process (95.08 million transistors/square mm). I'm going to randomly guess that the larger process means that they will be using at least 250% power, or more, to match the performance. And require more cooling (hence the fire icon).
65W 10 nm Intel SuperFin process Tiger Lake Core i3 11100B (2020-09-02) has 6 (12) Cores (threads), 480KiB L1, 7.5MiB L2, 12 MiB L3
65W 7FF TSMC process Ryzen 5 5600X (2020-11-05) has 6 (12) Cores (threads), 384 KiB L1, 3 MiB L2, 32 MiB L3
150W 12 nm TSMC process 3C5000L (2021) has 16 cores 64KiB Data and 64KiB instruction L1, 4MiB L2, 64 MiB L3
I'm also going to guess that all the caches will be larger (more cores, more power, more heat). But I do agree that it is a big deal for China, even if it is leaning on TSMC.
... to make customised versions of the Stuxnet worm to destroy Chinese chip foundries ... and once the US chip fabs are up, to infect and destroy TSMC's factory in Taiwan.
The West must win by any means possible - even mass extermination if required.
There is no more dangerous adversary in any space at the moment than China - they must be wiped out to the last person
*Cough cough* - I've taken my meds and am ok now - what did I miss?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
