In contrast to many computing traits, AI has sensible functions, so it has caught round past the preliminary fad interval. Persons are utilizing AI for something and all the things, even when it is not particularly-suited for the duty. (Word that every one of our articles are written by people!) If you wish to run giant AI fashions like GPT-4, you want extraordinarily giant and highly effective GPUs to get any type of respectable efficiency, and that mainly means shopping for NVIDIA {hardware}.
It is this unbelievable surge in demand for AI processing that has pushed NVIDIA’s inventory to the moon, giving the corporate a valuation within the one trillion US greenback vary, a quantity so giant it’s virtually inconceivable. The factor is, it might be even larger, besides TSMC cannot make NVIDIA’s GPUs quick sufficient.
Strictly talking, that is not fairly true, because it seems. TSMC could make chips very, in a short time. The issue is definitely with the superior packaging that NVIDIA’s Hopper GPUs require. You see, a part of the extraordinarily excessive efficiency of Hopper in AI duties is due to its use of unique HBM2 and HBM3 recollections. These require being put in on the identical bundle because the GPU chip itself, and this implies uncommon packaging.
Sadly, CoWoS can also be what’s holding up the manufacturing pipeline, it appears. Speaking to Nikkei AsiaTSMC Chairman Mark Liu stated that, opposite to well-liked perception, there is not any precise scarcity of AI chips. As a substitute, the corporate is cranking out CoWoS packages as quick as it will probably, and it is merely unable to maintain up with demand. In accordance with Liu, demand for CoWoS picked up “all of a sudden”, which apparently means it tripled in a yr.
Whereas being unable to serve your whole prospects’ demand is not nice, it is so much higher than having no demand in any respect. Liu says that he expects this bottleneck to be alleviated in “one and a half years.” That is as a result of enlargement of TSMC’s superior chip packaging capabilities with a brand new $2.9b USD facility particularly for these applied sciences.
What Liu did not say is how a lot further capability the brand new facility will add, however it’s most likely at the least double. We are saying that as a result of Liu acknowledged his perception that superior packaging like CoWoS is a part of a “paradigm shift” within the semiconductor business, the place as a substitute of constructing ever larger and denser processors, we as a substitute stack and mix a number of chips collectively into larger packages.