NVIDIA ensures more than 70% of TSMC’s cowos-l capacity to 2025

Nvidia closed business for more than 70% of TSMC Cowos-L Advanced Packaging Capacity for 2025driven by the strong demand for its GPUs based on architecture Blackwell.

Thus, the remittance volumes are expected to increase more than 20% in each quarterwith annual projections surpassing 2 million units this year alone.

TSMC Cowos-L Expansion

To meet the growing demand, the TSMC plans to expand its production facilities CoWoSincluding the construction of two new units in the Chiayi Scientific Park and two other acquired from Innolux.

However, the expected expansion for Chiayi Science Park Phase 2 was postponed to 2026redirecting efforts to build new factories in the STSP Fase III.

What is cowos-l and why is it vital to Nvidia?

O CoWoS-L (Chip-on-Wafer-on-Substrate – Large) is one of the advanced packaging technologies of the TSMC (Taiwan Semiconductor Manufacturing Company)designed to optimize the performance and efficiency of high performance chips.

How does cowos-L work?

O CoWoS It is a 2.5D packaging technology that allows the integration of multiple chips (or chiplets) e High Band Width Memory (HBM – High Bandwidth Memory) in a single silicon substrate. This improves the interconnection between components and reduces latency in communication between GPU and memory.

The version CoWoS-L differs by using UM redistribution layer (RDL) combined with a Ballows Silicon to interpose (LSI – Large silicon interposing).

This methodology significantly increases the size of the final chip, allowing a greater density of transistors and the stack HBMresulting in significant performance improvements and energy efficiency.

Difference between cowos-l, cowos-s and cowos-r

TSMC has already developed different variations of cowos, each with specific characteristics:

Why does NVIDIA bet on cowos-l?

O CoWoS-L It is necessary for Nvidia because it enables your line chips Blackwell (B200/B300) support More HBM Memoryincreasing bandwidth and processing capacity Large AI Models.

With this advanced packaging, Nvidia gets:

• Create GPUS larger and more powerfulwithout loss of thermal efficiency.
• Integrate more hbm modulesincreasing memory bandwidth.
• Improve energy efficiencyreducing consumption and heat generated.
• Increase production incomemaking the chips more viable economically.

Blackwell and the race for advancement in IA

As we saw above, the new line of chips B200/B300 of Blackwell architecture NVIDIA uses technology CoWoS-Lwhich allows Integration of more high bandwidth memory (HBM)increasing transistor density and optimizing performance for workloads Artificial intelligence and advanced computing.

A recent report pointed out that NVIDIA can consume up to 77% of global production of WAFERS aimed at AI processors in 2025. Such conjecture can characterize a wide mastery over critical components and thus limit the capacity of competitors as AMD e Intel to expand their own offerings in the AI ​​market.

Nvidia’s dominance over TSMC’s cowos-L production and technology reinforces its position as a leader in the artificial intelligence sector, making it increasingly difficult for competitors to dispute space in this rapidly growing market

Also read:

The search for new memory solutions

In addition to the dispute over wafers and packaging technology, Nvidia would also be evaluating the replacement of traditional DRAM memories with Future Products Socamm Modules.

Change can improve energy efficiency and optimize the use of space in graphic plates, which would be a differential for AI -centers and AI servers.

Therefore, the limited production capacity of the TSMC It can affect competitors and reinforce the industry’s dependence on the company, shaping the future of high performance computing in the coming years.

Fonte: Economic Daily News