Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/the-hbf-battle-is-underway-with-samsung-reportedly-joining-the-challenge-against-sk-hynix-and-kioxia.23750/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2030770
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

The HBF battle is underway, with Samsung reportedly joining the challenge against SK Hynix and Kioxia.

Fred Chen

Moderator
By Su Ziyun | Published October 2, 2025, 5:15 PM

According to Korean media reports, Samsung Electronics has begun the conceptual design and early development of high-bandwidth flash memory (HBF) products, preparing to join a new round of battle in the memory market.

HBF's design concept is similar to HBM, both using through-silicon via (TSV) technology to connect multiple layers of stacked chips. The difference is that HBM uses DRAM as its core, while HBF utilizes NAND flash memory, which offers higher capacity and more cost-effective features.

As data centers' demand for memory capacity increases, existing NAND capacity is becoming increasingly insufficient. Furthermore, HBM is volatile memory, losing all data upon power loss. Pairing it with non-volatile NAND Flash ensures data preservation while also meeting high-speed computing requirements. Therefore, the market is expected to gradually shift towards a dual-architecture model featuring HBM + HBF to meet the dual demands of high bandwidth and large capacity for AI computing power.

SK Hynix and SanDisk were the first to initiate HBF standard development and have signed a memorandum of understanding, aiming for sample delivery in 2026 and mass production in 2027. Meanwhile, Japan's Kioxia showcased its ultra-high-speed prototype in August of this year , demonstrating its technological prowess. These moves demonstrate that HBF has become a new battleground for global memory manufacturers, actively developing their presence.

Samsung Electronics , the leader in NAND flash memory market share, has officially entered the fray , seen as a key force potentially changing the market's trajectory. According to TrendForce data, Samsung's global NAND market share rose slightly to 32.9% in the second quarter , thanks to increased demand for enterprise-class SSDs driven by AI servers. Samsung maintains its top position. If this advantage is translated into the HBF market, it will make it difficult for other companies to enter the market.

As AI models rapidly expand in size , HBM alone may be unable to fully support the massive demands of data centers. Therefore, the industry is beginning to explore the potential of HBF as a complementary solution. If successfully implemented, HBF has the potential to provide GPUs with greater data capacity while ensuring low latency and reducing unnecessary costs.

 
Will there be a HBF standard soon? With CBA/wafer bonding you can do some interesting stuff with parallelism. But not sure how TSV (needed for parallelism with low latency) weighs in.
 
Will there be a HBF standard soon? With CBA/wafer bonding you can do some interesting stuff with parallelism. But not sure how TSV (needed for parallelism with low latency) weighs in.

Sandisk targets to deliver first samples of its HBF memory in the second half of calendar 2026 and expects samples of the first AI-inference devices with HBF to be available in early 2027.

 
Last edited:
As data centers' demand for memory capacity increases, existing NAND capacity is becoming increasingly insufficient. Furthermore, HBM is volatile memory, losing all data upon power loss. Pairing it with non-volatile NAND Flash ensures data preservation while also meeting high-speed computing requirements. Therefore, the market is expected to gradually shift towards a dual-architecture model featuring HBM + HBF to meet the dual demands of high bandwidth and large capacity for AI computing power.

"HBM is volatile memory, losing all data upon power loss. Pairing it with non-volatile NAND Flash ensures data preservation while also meeting high-speed computing requirements."

I’m not sure this will be a game-changing moment for data centers. Modern data centers are already highly reliable, equipped with uninterruptible power supply (UPS) systems, redundant power sources, battery backup systems (even down to the board level), and high-performance, fault-tolerant storage systems.

HBF might offer certain advantages for data center operations, but I doubt the additional cost is justifiable for preventing data loss caused by a power outage.
 
"HBM is volatile memory, losing all data upon power loss. Pairing it with non-volatile NAND Flash ensures data preservation while also meeting high-speed computing requirements."

I’m not sure this will be a game-changing moment for data centers. Modern data centers are already highly reliable, equipped with uninterruptible power supply (UPS) systems, redundant power sources, battery backup systems (even down to the board level), and high-performance, fault-tolerant storage systems.

HBF might offer certain advantages for data center operations, but I doubt the additional cost is justifiable for preventing data loss caused by a power outage.
Agreed, the power outage is not a driving reason.

Probably the main driver is the much larger capacity, and also savings from no refresh.

The endurance and latency are still drawbacks, so HBM cannot be completely replaced.
 
Agreed, the power outage is not a driving reason.

Probably the main driver is the much larger capacity, and also savings from no refresh.

The endurance and latency are still drawbacks, so HBM cannot be completely replaced.
That’s true. I perceive the main benefit for HBF is to be able to run trillions parameter models for inference on a single Nvidia chip. I doubt it would assist significantly with training.
 
Back
Top