Latest news

NVIDIA capitalizes, but TSMC controls production of AI processors

While the whole world admires Nvidia for its market capitalization of over $1 trillion at the beginning of the year and praises the company for its prominent positioning in the megatrends of artificial intelligence (AI) and high-performance computing (HPC), there is another company that both benefits significantly from these megatrends and has almost complete control over the production of AI processors. TSMC, according to a report by DigiTimes, is the company making some of the most complex processors for AI and HPC machines ever developed.

Source: ServeTheHome

Tens of billions of dollars

Customers include well-known companies such as NVIDIA, AMD, Intel, Tenstorrent, Cerebras and Graphcore, to name a few. TSMC manufactures Nvidia’s popular A100 and H100 compute GPUs for AI and HPC workloads, as well as their A800 and H800 derivatives for the Chinese market. TSMC also produces AMD’s EPYC CPUs and Instinct GPUs. Even emerging players in AI and HPC like Tenstorrent and developers of innovative products like Cerebras with their wafer-scale processors have chosen TSMC as their manufacturing partner.

Although TSMC does not publish specific figures on revenue from sales of CPUs, GPUs and specialized processors or SoCs for AI, data centers, HPC and servers, it can be assumed that the company generates tens of billions of dollars due to the high silicon requirements of these products. One example is NVIDIA’s GH100 compute GPU with a die size of 814 mm2, while AMD’s EPYC “Genoa” uses 12 Zen-4-based CCD chips, each of which is about 72 mm2.

In total, this requires an area of 864 mm2 of N5 silicon. Although exact revenue breakdowns for TSMC’s competitors, Samsung Foundry and GlobalFoundries, are not available, it is safe to say that TSMC benefits from AI and HPC in general. Compared to these companies, TSMC is clearly ahead as the leading contract chip manufacturer. In particular, TSMC dominates in the supply of AI GPUs, supplying both NVIDIA (which controls over 90 er shipments) and AMD (which controls less than 10%).

AI and HPC are gaining importance for TSMC

TSMC provides a fairly detailed breakdown of its revenue, making clear distinctions between automotive, IoT, smartphones, and high performance computing (HPC). However, this breakdown is not detailed enough to distinguish between chips for AI, HPC, client PCs, servers, and game consoles. Within TSMC, all of these processors and system-on-chips (SoCs) fall into the HPC segment, which is currently thriving.

In 2019, HPC products accounted for 30 of TSMC’s revenue, or $10.389 billion. In the same year, smartphone SoCs accounted for 49 es of TSMC revenue, equivalent to $16.97 billion. However, the share of HPC products in TSMC revenue continues to increase: 33% in 2020 ($15 billion), 37% in 2021 ($21 billion), and 41% in 2022 ($31.11 billion). For smartphone SoCs, on the other hand, the trend is reversed: they will account for 39 es TSMC sales in 2022 (USD 29.59 billion).

While AMD and NVIDIA source large volumes of data center-focused chips from TSMC, Apple remains the largest customer of the world’s leading chipmaker. This is especially true now that the company manufactures both smartphone and PC SoCs, which fall into the HPC category. According to DigiTimes, Apple alone will be responsible for about 23 es of TSMC’s total revenue in 2022.

More chips on the way

The semiconductor industry is experiencing a recovery after a period of downturn, and the growing focus on generative AI is creating new momentum in the market. Nvidia is benefiting from this upswing in AI with its A100/A30/A800 and H100/H800 GPUs, which are manufactured at TSMC. Similarly, AMD is ramping up its orders with TSMC for its upcoming Instinct MI300 series products, which will be mass-produced on TSMC’s N5 node starting in the second half of 2023. According to a report from DigiTimes, Apple, AMD and NVIDIA have also committed to manufacturing future chips using TSMC’s N3 (3nm-class) and N2 (2nm-class) manufacturing technologies.

Source: TomsHardware



Lade neue Kommentare



1,432 Kommentare 787 Likes

TSMC und Apple: Hier darf man auch nicht vergessen, daß Apple bei jedem neuen Fertigungsknoten, den TSMC für Großserienfertigung
auflegt, wohl auch vertragsmässig das "right of first refusal" hat. Die wenigen Male, als Apple das nicht genutzt hat, warren, als es nicht rechtzeitig für die nächste iPhone/Bionic Generation fertig war.
Große Chips für KI usw werden auch eher in etablierten Knoten gefertigt (geringere Fehler- und Ausschussquoten), was bei TSMC 5 und 7 nm sind, und jetzt auch 4 nm. 3 nm ist erst mal für Smartphone SoCs.
Und die Abhängigkeit bei KI und Beschleunigern von Fertigung bei TSMC sind auch ein Hauptgrund, warum die USA und jetzt auch Europa (Magdeburg!) massiv in EUV Foundries investiert haben. Bei den USA kommt noch die Notwendigkeit dazu, Chips von "trusted Foundries" beziehen zu müssen, wenn die Rechenanlagen irgendwas mit "National Security" zu tun haben. Auch daher sind die Subventionen für Intels EUV Foundries in den USA durchaus Teil der Verteidigungsausgaben, zumindest im weiteren Sinn.

Antwort Gefällt mir



1,348 Kommentare 550 Likes


Apple hat "right of first refusal", und das nur wenige male nicht genutzt. => ergibt für mich dass sie fast immer "refusen".

Wer von uns hat da jetzt den Knoten, wenn dieses Recht eigentlich heißt, dass Apple als Erstes gefragt werden muss ob und wieviel sie von den neuen Kapazitäten wollen; wobei sie scheinbar fast immer was wollten.

Antwort Gefällt mir

Danke für die Spende

Du fandest, der Beitrag war interessant und möchtest uns unterstützen? Klasse!

Hier erfährst Du, wie: Hier spenden.

Hier kannst Du per PayPal spenden.

About the author

Igor Wallossek

Editor-in-chief and name-giver of igor'sLAB as the content successor of Tom's Hardware Germany, whose license was returned in June 2019 in order to better meet the qualitative demands of web content and challenges of new media such as YouTube with its own channel.

Computer nerd since 1983, audio freak since 1979 and pretty much open to anything with a plug or battery for over 50 years.

Follow Igor:
YouTube Facebook Instagram Twitter