I already said as much. Training is way more computationally expensive than inference. It may take huge resources and huge amounts of tests to make a proper AI model for something. But way less resources would be needed to use that model in real world situations. Those training nodes might be huge centralized supercomputers chewing huge datasets. What does it matter for such strategic electronics if they take 2x or even 4x the power and space?The policy is a failure even for the narrow objectives they've defined. Advances in AI, etc. are much more driven by architectures (both hardware and software) and algorithms, not feature sizes. China is able to both produce 14nm/7nm indigenously (with ASML equipment for now and SMEE equipment shortly) and import more advanced chips. Servers are built with readily available commodity chips, and even banned companies can easily set up shells and source them.
The electronics people will actually use in the field, with the models baked in, will be portable and cheap.
This means having the transistor advantage won't be as important for military applications as you would think. And even if it is, if training can be shrunk to a field portable apparatus, if a human with automatic assists turns out to be cheaper than that system, it will still be of no advantage. I think the advantage will only come into play a long time into the future. Not in less than a decade if ever. The brain is 3D. And neurons don't store 1s and 0s. The brain operates at like 100x slower speed than semiconductors do. But once you add the fact it is non-binary the difference is eliminated and with the 3D the brain wins easily.
Gosh I hope so. But I would not bet on it. I think China would be better off making immersion DUV with 450mm wafers. Kill the West with volume of legacy nodes to a price point they can't compete.Medium term (3 to 5 years) China will have completely indigenized a cutting edge semiconductor design and manufacturing chain centered around domestic EUV lithography.
I said this much here some time ago. We discussed here how HiSilicon had made its own Gallium Arsenide RF modules. But guess what. They were fabbed in Taiwan at WIN Semiconductors. That factory seems to be the only non-US competition in the commercial market.I have some questions that are tangentially related to semiconductors:
1) in this report it is said that high-end radio frequency (RF) components are monopolised by the usa; "U.S. companies Skyworks, Qorvo, and Broadcom “monopolize” the high-end market for the “key” RF component, the power amplifier chip (or RF chip). There are “essentially” no Chinese firms at all in the high-end RF chip market. U.S. manufacturers such as Qorvo “entirely” dominate the market for another class of RF components: high-end wave filters used in mobile phones.
if china improves its semiconductor industry, is this one of the sectors where over reliance on the usa could be lessened?
You could call them the TSMC of Gallium Arsenide chips, while Qorvo is a vertically integrated company sort of like Intel.
I am sure the Chinese MIC i.e. CETC has its own facilities. GaA RF is used in military radar. But AFAIK there are no commercial facilities in China. However I would not be surprised if you could use Gallium Nitride for much the same purposes. And there are loads of factories with Gallium Nitride process in China. It is commonly used to make consumer LEDs and modern military radar or 5G RF.
For example this is a Chinese manufacturer of substrates and discrete GaN LEDs and transistors.
Qorvo and other GaA RF makers in the US work for the DoD you are correct. The Chinese MIC has its own capabilities. It just seems they don't share with the commercial sector. China needs to stop this.2) I think I read somewhere here on the forum that american aesa radars have qorvo chips in them. and I've learnt that chinese military aesa radars are very good quality. If point 1 is true, how did china manage to achieve such quality for aesa radars if they rely on american chips that are probably sanctioned by now? do chinese companies use home-grown" chips that are from bigger nodes but that the chinese semiconductor industry has already mastered?
Last edited: