Apple has been using TSMC’s N5 nodes nearly exclusively for its system-on-chips for smartphones and PCs for the past few years. TSMC has had to improve its production capacity as more companies use these fabrication processes. According to a fresh source, TSMC will raise its N5 production capacity by roughly 25% this year to accommodate demand from companies like AMD, Nvidia, and MediaTek.
Vanilla N5, performance-enhanced N5P, N4, N4P, N4X, and Nvidia-specific 4N are all part of TSMC’s N5 (5nm-class) manufacturing process family. Apple is thought to be using N5 and N5P for its current A14, M1, and A15 system-on-chips, although AMD, MediaTek, and Nvidia are expected to utilise a variety of technologies from the lineup. Meanwhile, Apple’s next-generation A16 processor is expected to switch to N4 as well.
Nvidia, for example, has chosen 4N for its Hopper compute GPUs (and maybe for Ada Lovelace consumer GPUs), whereas MediaTek has chosen N5 for its Dimensity 8000/8100 and will employ N4 for the Dimensity 9000.
According to a DigiTimes article, TSMC can produce up to 120,000 wafer starts per month (WSPM) with N5. By early 2022, TSMC had expected to produce 120,000 N5 WSPM, indicating that the foundry has reached its full capacity. TSMC will install extra equipment to raise N5 output to 150,000 WSPM by Q3 2022 to better service existing and future clients interested in one of the N5 processes. By mid-2022, TSMC will require more N5-capable equipment at its fabs.
Given the length of modern cycles, we’re pretty sure TSMC is already ramping H100 production using the Nvidia-tailored N4 node. Nvidia plans to start shipping it’s Hopper compute GPUs commercially in Q3, so given the length of modern cycles, we’re pretty sure TSMC is already ramping H100 production using the Nvidia-tailored N4 node. While the production volumes dedicated to these GPUs aren’t particularly enormous, the chips themselves are quite large, consuming a major portion of TSMC’s N5-capable capacity.
Meanwhile, because Apple typically ramps up manufacturing of new iPhone SoCs around April or May, TSMC is expected to start producing the A16 in the coming weeks. Because Apple’s smartphone SoCs are utilised in hundreds of millions of devices, the company will continue to be TSMC’s largest customer in terms of revenue and processed wafers. Also, because MediaTek sells a lot of complex SoCs these days, it will need tens of millions of TSMC’s Dimensity 8000/8100/9000 application processors and will thus remain the company’s second-largest customer.
This fall, AMD is also expected to unveil its next-generation Zen 4-based Epyc and Ryzen CPUs, as well as RDNA 3-based Radeon RX 7000-series GPUs. Around the same time, Nvidia will unveil its Ada Lovelace-powered GeForce RTX 40-series consumer solutions. Because all of these items are intended to be widely available, TSMC will require a large amount of capacity to manufacture them.
Both AMD and Nvidia have spent billions of dollars at TSMC to ensure that they can receive all of the processors they require. As a result, TSMC’s new capacity has practically been paid for months before it becomes operational.