NVIDIA Reportedly Overwhelms TSMC with 3 and 4-Nanometer Orders

March 5, 2024 -- The annual AI event, NVIDIA GTC (GPU Technology Conference), is set to take place on March 17th, as H200 and the next-generation B100 will reportedly be announced ahead of schedule to seize the market. According to Commercial Times’ report citing sources, H200 and the upcoming B100 will adopt TSMC’s 4-nanometer and 3-nanometer processes respectively. H200 is expected to be launched in the second quarter, while it’s rumored that orders for the B100 adopting Chiplet architecture have already been placed for production.

Sources cited by the report also indicate that NVIDIA’s orders are robust, pushing TSMC’s 3 and 4-nanometer production capacity to near full utilization, making the first quarter, traditionally a slow season, unexpectedly busy.

Regarding the matter of NVIDIA’s next-generation chip orders overwhelming TSMC’s advanced processes, TSMC stated that details regarding production capacity remain consistent with the previous earnings call and will not be elaborated further.

Still, Commercial Times further cited industry sources, revealing that TSMC, in response to anticipated capacity constraints by 2023, is accelerating its efforts. Particularly focusing on advanced packaging like CoWoS, they’ve not only relocated equipment from the Longtan facility but also swiftly activated the AP6 plant in Zhunan.

Another industry sources reportedly indicate that the planned construction of the Tongluo
facility, initially slated for the second half of this year, is now scheduled to commence in the second quarter. The aim is to ramp up 3D Fabric capacity to produce 110,000 12-inch wafers per month by the first half of 2027.

Meanwhile, TSMC’s advanced processes remain fully utilized, with capacity utilization exceeding 90% in February, driven by sustained AI demand.

NVIDIA, on the other hand, recently emphasized that computational-intensive tasks like Generative AI and large language models require multiple GPUs. From customer purchase to model deployment, it takes several quarters. Thus, this year’s inference applications stem from GPU purchases made last year. As model parameters grow, GPU demand is expected to expand.

In addition to increasing GPU quantities, NVIDIA’s GPU efficiency is poised for a significant boost this year. The Blackwell series, notably the B100, is hailed as NVIDIA’s next-generation GPU powerhouse by the market.

Not only is it the first to adopt TSMC’s 3-nanometer process, but it’s also the pioneer in Chiplet and CoWoS-L packaging among NVIDIA products. This tackles high power consumption and cooling issues, with projected single-card efficiency and transistor density expected to surpass AMD’s MI300 series set to debut in the first quarter.