AIMemoryNewsNVIDIAPC Hardware

Report Says TSMC Began Drafting 3D DRAM Plans for Nvidia. TSMC Says It’s a Logic Foundry.

Nvidia can’t get enough memory, and neither can anyone else. Now a report says Nvidia asked the one company that already makes its GPUs to start making memory too. TSMC’s answer, so far, is that it builds logic chips and plans to keep doing that.

What was reported

On Friday, November 6, The Information reported that TSMC recently began drafting plans to make 3D DRAM at the request of Nvidia, its biggest customer. The outlet cited two people with direct knowledge of the plans. Reuters relayed the report, and so did Bloomberg. Both attributed it to The Information.

According to that reporting:

  • TSMC could build new factories for memory, or buy an existing plant and convert it, the way Micron did this year.
  • The timing of any production is unclear.
  • The work is led by Hsu Kuo-chin, the TSMC vice president who oversees advanced packaging. Tom’s Hardware notes he used to run Micron’s operations in Taiwan.

3D DRAM stacks layers of memory cells on top of each other inside the chip, instead of laying them out on one flat plane. More layers means more capacity per wafer. That is different from HBM, which stacks finished DRAM dies on top of each other and sits next to the GPU.

What TSMC says

TSMC and Nvidia did not immediately respond to Reuters. TSMC did answer Tom’s Hardware. A spokesperson said the company “focuses on our core business at all times, which is dedicated IC foundry,” and that its mission is to be the trusted provider for the “global logic IC industry for years to come.”

Tom’s Hardware reads that as a denial. To be precise, the statement doesn’t address the report’s specific claims. It doesn’t say whether plans were drafted, or whether Nvidia asked. Nvidia hasn’t commented publicly.

Why Nvidia would ask

AI data centers are soaking up DRAM, and the shortage has pushed memory prices up across the industry. Only three companies make leading-edge DRAM at scale: Samsung, SK hynix and Micron. A fourth supplier, especially one Nvidia already works with on every flagship GPU, would give it more leverage.

It also fits a pattern. In October, Reuters reported, citing the Seoul Economic Daily, that Nvidia asked all three memory makers to build the base dies of its custom HBM on TSMC’s 3nm process. Nvidia has been pulling TSMC closer to its memory for months.

Why this won’t fix RAM prices soon

Even if the report is right, don’t expect cheaper RAM next year. Tom’s Hardware’s analysis makes three points:

  • TSMC has world-class logic processes but no competitive DRAM process of its own. Memory needs very different manufacturing.
  • 3D DRAM is still years away even for the big three. Micron plans at least four more generations of conventional flat DRAM before it gets there.
  • For those reasons, the outlet says a 3D DRAM push can’t ease the shortage for at least five years.

Memory is also a boom-and-bust business. Shortages bring new factories, and new factories bring gluts. That risk is a good reason for a company as profitable as TSMC to stay careful.

PCGeek’s take

This is a report, not a plan TSMC has announced, and TSMC’s public line is that it stays a logic foundry. Both things can be true at once. Big chipmakers draft plans all the time that never become fabs. What the report does show is how worried Nvidia is about memory supply. When the biggest name in AI chips is reportedly asking its foundry to get into DRAM, the memory crunch that is raising prices on GPUs, laptops and RAM kits isn’t ending soon.

We’ll update this if TSMC or Nvidia says more.

Photo: TSMC Fab 12B in Hsinchu, Taiwan, by 曾 成訓 via Wikimedia Commons, CC BY 2.0. It is a TSMC logic fab, not a memory plant.

Would you trust TSMC-made memory in your next GPU, or should it stick to logic? Tell us in the comments.

Follow @pcgeekinfo for more.

PCgeek

Techie, YouTuber, Writer, Creator

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.