Micron Sold All of Its High Bandwidth Memory Supply for 2024 and Most of 2025

[Daniel Nenni]

The high-bandwidth memory market is exploding due to its usage in AI accelerators.

Credit: Micron

We all know Nvidia is enjoying life as the belle of the AI ball, thanks to its hardware being the gold standard for training AI models. Now, it appears it'll be bringing its hardware partners along for the ride as well. Memory maker Micron has been tapped to provide the latest high-bandwidth memory (HBM3e) for the company's upcoming H200 accelerator, and it says it's completely sold out for the rest of 2024. Not only that, but it also says most of its HBM supply is already allocated for 2025.

Micron executives recently shared the results of a successful quarter for its memory products, which include HBM for servers, NAND for SSDs, and DDR5 for clients. The quarterly earnings statement is jubilant, with CEO Sanjay Mehrotra calling the results "well above the high end of guidance." Turning to the company's HBM memory portfolio, Mehrotra said, "Our HBM is sold out for calendar 2024, and the overwhelming majority of our 2025 supply has already been allocated." Due to the company's juicy contract with Nvidia for HBM3e chips for its H200 accelerator, which will begin shipping this year, the company expects business to be rosy for some time.

Micron is expected to roll out 12-layer stacks of HBM3e in 2025, eventually pushing bandwidth to 2TB/s and beyond in 2026. Credit: Micron

Micron says its HBM3e memory, the fastest server memory available until HBM4 arrives in 2026, has 30% lower power consumption than its competitors' offerings. It competes with both Samsung and SK Hynix in this market and is seen as the underdog, given its small market share. However, its status is on the upswing due to its partnership with Nvidia and what appears to be a competitive memory portfolio. Its HBM3e memory "cubes" offer 24GB capacity in an eight-layer design for up to 1.2TB/s of bandwidth per cube. The H200 will feature six modules for 141GB of HBM3e memory.

Tom's Hardware notes Micron beat its rivals to the HBM3e market, which likely granted it the coveted position of being added to Nvidia's second-generation H200 product. The company is readying its second salvo into the HBM3e market with a 12-layer design, increasing capacity per-cube to 36GB. In the earnings call, Mehrotra said Micron expects to begin ramping this version throughout 2025, so the eight-layer design will be its money-maker in 2024 as it rides along the gravy train that is the Nvidia H200 accelerator.

https://www.extremetech.com/computing/micron-sold-all-of-its-high-bandwidth-memory-supply-for-2024-and-most-of

 

[Daniel Nenni]

GO AI!!!!!! I was hoping Micron would get big CHIPS Act money. We need more US born memory, absolutely.

 

[nghanayem]

I think the more impressive thing went under most folks radar:

We continue to mature our production capability with extreme ultraviolet lithography and have achieved equivalent yield and quality on our 1-alpha as well as 1-beta nodes between EUV and non-EUV flows. We have begun 1-gamma DRAM pilot production using EUV and are on track for volume production in calendar 2025. The development of our next-generation NAND node is on track, with volume production planned for calendar 2025. We expect to maintain our technology leadership in NAND.

To my knowledge Micron has never once mentioned 1-alpha/beta having an EUV versions. They even went out of their way talking about how they are leading just with advanced multipatterning. Micron sticking with optical and pulling ahead of Hynix and especially Samsung who were more aggressive on EUV (Samsung inserting it back on their 1Z node). The question was if Micron would stumble on 1-gamma like the Koreans did on their first EUV implementations? Obviously the tools and ecosystem in 2024 are much better than they were in 2020, but it is still a big change that Micron would have needed to adapt to casting some uncertainty on Micron. The fact that Micron was just hiding this ace up their sleeve, I feel erases almost all doubts about if the 1-gamma transition would be a graceful one. It really does feel like a mic drop moment vs their larger/richer Korean rivals that "Hey we can beat you only using optical; and oh by the way we also figured out EUV in parallel so any technological edge you thought you had doesn't even exist.".

 

[hist78]

A short documentary talked about the Micron Taiwan.

 

[Fred Chen]

I think the more impressive thing went under most folks radar:

To my knowledge Micron has never once mentioned 1-alpha/beta having an EUV versions. They even went out of their way talking about how they are leading just with advanced multipatterning. Micron sticking with optical and pulling ahead of Hynix and especially Samsung who were more aggressive on EUV (Samsung inserting it back on their 1Z node). The question was if Micron would stumble on 1-gamma like the Koreans did on their first EUV implementations? Obviously the tools and ecosystem in 2024 are much better than they were in 2020, but it is still a big change that Micron would have needed to adapt to casting some uncertainty on Micron. The fact that Micron was just hiding this ace up their sleeve, I feel erases almost all doubts about if the 1-gamma transition would be a graceful one. It really does feel like a mic drop moment vs their larger/richer Korean rivals that "Hey we can beat you only using optical; and oh by the way we also figured out EUV in parallel so any technological edge you thought you had doesn't even exist.".

DRAM has established layout styles and much lower mask count than foundry. So there is really no driving factor for EUV there. Some layout patterns are actually optically incompatible on EUV (due to X/Y differences).

 

[Fred Chen]

2025 timing makes me wonder about Micron's 3D DRAM plans..