Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/samsung-1a-dram-re-design-considered-for-hbm-competitiveness.21262/
        )

    [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] => 2021770
            [XFI] => 1050270
        )

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

Samsung 1a DRAM re-design considered for HBM competitiveness

Fred Chen

Moderator

Samsung, 1a DRAM redesign in jeopardy… HBM competitiveness recovery 'super strong'​

[Samsung DRAM's Tasks abc ①] Developing Countermeasures to HBM Sluggishness… Outlines to be Outlined by the End of the Year

Semiconductor/Display Input: 2024/10/15 16:55 Modified: 2024/10/16 14:38

Reporter Jang Kyung-yoon

The crisis surrounding Samsung Electronics is raising its head again. The source of the crisis seems to be largely centered on semiconductors. Among them, the recovery of memory competitiveness centered on the HBM business and the sluggish foundry are being pointed out. The rebuilding of the organizational culture of trust and communication is also key. Accordingly, ZDNet Korea will examine the fundamental problems of the Samsung crisis theory and seek opportunities for a leap forward together. [Editor's Note]

Samsung Electronics is in crisis. All of its business divisions, including semiconductors, home appliances, MX, and SDC, are struggling, but it is especially painful that its main business, memory, failed to shine even in the boom cycle of the third quarter of this year.

Among them, the market’s disappointment in the HBM (high bandwidth memory) business was great. Samsung Electronics had been working behind the scenes to start supplying HBM3E to Nvidia in earnest from the third quarter of this year. However, it has not yet passed the quality test for the 8-layer product, and there is a high possibility that the 12-layer product will be delayed until the second or third quarter of next year.

There are complex reasons for Samsung Electronics' delay in commercializing HBM. However, experts point out that the fundamental problem with HBM is the problem of the core die, DRAM. In the HBM structure where multiple DRAMs are stacked vertically and connected, the performance of DRAM is inevitably directly linked to the performance of HBM.

EUV preemptive application, but… 1a DRAM competitiveness shaken

From this perspective, the point at which Samsung Electronics' position as number one in DRAM technology was greatly shaken is pointed out to be the '1a DRAM'.

10nm-class DRAM has evolved in the order of 1x (1st generation) - 1y (2nd generation) - 1z (3rd generation) - 1a (4th generation) - 1b (5th generation). 1a DRAM has a line width of 14 nanometers (nm). Samsung Electronics began mass production in the second half of 2021.

Samsung Electronics was unable to mass-produce 1a DRAMs faster than its competitors, but it sought to increase its competitiveness by more actively introducing cutting-edge technologies such as EUV (extreme ultraviolet lithography). The number of EUV layers applied by Samsung Electronics to 1a DRAMs was 5, which was more than that of its competitor SK Hynix (1).

However, the prevailing opinion is that such attempts have not led to successful results as of now. EUV is advantageous in miniaturizing line widths compared to ArF (argon fluoride), the existing exposure process (the process of engraving circuits on semiconductors). Therefore, the advantage of EUV is that it can increase process efficiency and reduce manufacturing costs, which are the core of memory.

However, EUV has high technical difficulty, which has been a factor in reducing process stability during actual mass production. As a result, the cost of Samsung Electronics' 1a DRAM did not decrease as initially expected.

The DRAM design itself is also evaluated as not being perfect. In particular, it is known that the development of server products was delayed compared to competitors due to setbacks. In fact, SK Hynix was the first to receive certification for server DDR5 products based on 1a DRAM from Intel in January of last year.

Discussion of ‘redesign’ amid delay in HBM commercialization… Attempt at a major transformation

There are also claims that the performance of 1a DRAM is holding back Samsung Electronics' recent sluggish mass production of HBM3E for Nvidia.

Recently, Samsung Electronics conducted a field test with NVIDIA on the HBM3E 8-stage product at the Pyeongtaek campus. NVIDIA concluded the field test itself without any particular issues. However, it is known that they evaluated that the data processing speed of HBM is lower than that of other products.

According to a comprehensive analysis of the stories inside and outside of Samsung Electronics, Samsung Electronics’ HBM3E 8-stage data processing speed (Gbps) is about 10% lower than that of SK Hynix and Micron. The specific figures vary depending on test results and customers, but there is no disagreement that the performance is lacking compared to the two competitors that use 1b DRAM.

Accordingly, Samsung Electronics is considering a 'super-strong' measure to restore the fundamental competitiveness of server DRAM and HBM under the leadership of Vice Chairman Jeon Young-hyun.

In a recent apology following the announcement of the provisional third quarter results, the former vice chairman stated, “Above all, I will restore our fundamental technological competitiveness,” and “Technology and quality are our lifeblood and Samsung Electronics’ pride that we can never compromise on.” He also said, “Rather than short-term solutions, I will secure fundamental competitiveness,” and “Furthermore, I believe that the only way for Samsung Electronics to make a comeback is through new technologies that have never existed in the world and perfect quality competitiveness.”

According to multiple sources familiar with the matter, Samsung Electronics has recently been internally discussing a plan to redesign some of the circuitry of its 1a DRAM.

An industry insider explained, "Samsung Electronics, which is struggling with its memory strategy, has concluded that it needs to redesign its 1a DRAM," adding, "However, the final decision has not been made, and it is a situation where bold decisions are needed because it will require taking on many risks."

 
Wow, what a huge problem. Micron held off inserting EUV all these years, until their upcoming node for 2025, avoiding all of these issues.
Normally, the HBM issue would sound like a TSV origin, but it seemed the performance of server DDR5 was also afflicted, so they suspected the 1a design.

So it is also surprising that SK Hynix and Micron have been able to proceed to 1b without the issues. Yet even for the 1b (12 nm) designs there's no strong reason for EUV.
 
Normally, the HBM issue would sound like a TSV origin, but it seemed the performance of server DDR5 was also afflicted, so they suspected the 1a design.

So it is also surprising that SK Hynix and Micron have been able to proceed to 1b without the issues. Yet even for the 1b (12 nm) designs there's no strong reason for EUV.
Agreed, most HBM issues are TSV related and Samsung suffers with that too.

But in this case I think we are talking about a poor D5 1a core design from Samsung from the start. Overclockers understood very early that Samsung D5 couldn't be pushed much beyond the stated specs. In addition you only need to look at Samsung vs Hynix relative D5 spot prices to realise there is something amiss in the wider market too.

Too me, this is all downstream of Samsung Semi's reversal in sucess & culture which has been going on for nearly 10 years now but which has only really become apparent in the past 2~3 years. D5/HBM is not an isolated issue, please see 3D NAND and Foundry too.

There are some [very broad] similarities to Intel here, although not as severe. Will Samsung be able to get back? Yes, I think they will. Will they be able to re-assert their prior dominant position? I doubt it.
 
Too me, this is all downstream of Samsung Semi's reversal in sucess & culture which has been going on for nearly 10 years now but which has only really become apparent in

There are some [very broad] similarities to Intel here, although not as severe. Will Samsung be able to get back? Yes, I think they will. Will they be able to re-assert their prior dominant position? I doubt it.
Your sense is that Intel’s problems are far deeper than Samsung’s issues? Just wanted to confirm your comment.
 
Your sense is that Intel’s problems are far deeper than Samsung’s issues? Just wanted to confirm your comment.
Yes. They both have culture and leadership problems but going forward at least memory retains its relative value, whereas CPU does not.
 
- Micron managed to develop 1b node without EUV, and managed to produce 1b-based HBM
- SK hynix successfully migrated EUV and mass producing 1b-based HBM
- Samsung is struggling to stablize yield of 1b node (EUV), no plans for 1b-HBM
- SK hynix is gaining tons of market share, Samsung flat and micron is losing A LOT

I think this is one of the most interesting observation in semiconductor industry. Juding from massive market share gain of SK Hynix and lose of Micron, it seems that 1a~1b node was the right insertion window of EUV. So in a sense, Samsung was also right but they seriously dropped the ball so they were not able to utilize EUV experience from foundry.

I think Samsung is somewhat suffering from their old mindsets. Maybe 1a and 1b were too cost-oriented so they have no rooms to use it for derivative products(HBM, for example). This is somewhat understandable since they also suffered 2021~2022 memory downturn. But 1a redesign NOW means that they're not even communicating inside(apart from their transparency from outside), since 1a-based HBM was in NVIDIA qualification for a year, enough to revise their 1a HBM derivative if something's going wrong.
 
- Micron managed to develop 1b node without EUV, and managed to produce 1b-based HBM
- SK hynix successfully migrated EUV and mass producing 1b-based HBM
- Samsung is struggling to stablize yield of 1b node (EUV), no plans for 1b-HBM
Let's not forget Micron was first with 1b, without sacrificing bit density.
 
- Micron managed to develop 1b node without EUV, and managed to produce 1b-based HBM
- SK hynix successfully migrated EUV and mass producing 1b-based HBM
- Samsung is struggling to stablize yield of 1b node (EUV), no plans for 1b-HBM
- SK hynix is gaining tons of market share, Samsung flat and micron is losing A LOT

I think this is one of the most interesting observation in semiconductor industry. Juding from massive market share gain of SK Hynix and lose of Micron, it seems that 1a~1b node was the right insertion window of EUV. So in a sense, Samsung was also right but they seriously dropped the ball so they were not able to utilize EUV experience from foundry.

I think Samsung is somewhat suffering from their old mindsets. Maybe 1a and 1b were too cost-oriented so they have no rooms to use it for derivative products(HBM, for example). This is somewhat understandable since they also suffered 2021~2022 memory downturn. But 1a redesign NOW means that they're not even communicating inside(apart from their transparency from outside), since 1a-based HBM was in NVIDIA qualification for a year, enough to revise their 1a HBM derivative if something's going wrong.
Why do you believe micron is losing a lot ?
 
Let's not forget Micron was first with 1b, without sacrificing bit density.

Why do you believe micron is losing a lot ?

Micron released world 1st 1b node in late 2022. So it's already 2 year product, it should be on HVM

- SK Hynix got +32% memory(DRAM+NAND) revenue increase QoQ (Q1 ~ Q2), DRAM +20% bit shipment increased
- Samsung got memory +24% memory(DRAM+NAND) revenue increase QoQ (Q1 ~ Q2), DRAM bit shipment change unknown
- Micron got +17% memory(DRAM+NAND) revenue increase QoQ (FQ2 ~ FQ3, feb ~ may), DRAM bit shipment decreased (why?)

This is what I think is interesting. Between Q1~Q2, Samsung didn't get HBM3E qualification from NVIDIA, so we could assume that this difference mostly comes from 1a~1b core products. Of course, they both sell NAND as well but NAND market is rather weak now and Samsung is suffering more from NAND than others. AFAIK, Micron 1b is more than decent product in terms of power and area(maybe the best in 2023~2024) but somehow it doesn't seem materializing in the market yet.
 
That's easy. The ghost of Christmas past is haunting Micron. They are the smallest major DRAM and smallest major flash manufacturers. The memory industry is all about running flat out and maximizing economics of scale. The cost per bit advantage of say 1-alpha to 1-beta is much smaller than say 4MB vs 16MB DRAM. While Micron isn't subscale their process lead (and likely higher yield since they have a TTM lead) probably only just about makes up for the scale differences. There is also the reality that Micron (and also Kioxia) has no "big brother" to ensure they are well funded. Samsung has the might of Samsung electronics and Samsung group at their back, while SK-Hynix has SK group at their back.
 
I think we are going to see a paradigm shift in the DRAM sector with 3D-DRAM replacing the current planar memory.
That will enable huge advances in bit density like we saw with V-NAND.
 
Micron is years into HVM of 1b. Interestingly, they have been concurrently producing 1b and 1a using EUV to evaluate when to adopt EUV, and have claimed doing so with similar yield.
 
I think we are going to see a paradigm shift in the DRAM sector with 3D-DRAM replacing the current planar memory.
That will enable huge advances in bit density like we saw with V-NAND.
The hold-up seems to be, unlike 2D DRAM or 3D NAND, the basic architecture is still not settled. Maybe 4F2 is easier to be consistent on, though that might last only a few nodes.
 
Maybe some, but from the 2022 evaluation by TechInsights, even when 1a was the leading edge, HBM still had 1y, 1z, so less leading edge still applies. SK Hynix is essentially the HBM pioneer (first chip in 2013) so it has the most advantage.
The foundation of Micron’s high-bandwidth memory (HBM) solution is Micron’s industry-leading 1β (1-beta) DRAM process node
 
Maybe some, but from the 2022 evaluation by TechInsights, even when 1a was the leading edge, HBM still had 1y, 1z, so less leading edge still applies. SK Hynix is essentially the HBM pioneer (first chip in 2013) so it has the most advantage.
I'm sorry this was not HBM but simply DDR5 comparison.

HBM3E is the most recent HBM to come out: https://www.trendforce.com/presscenter/news/20240930-12319.html/ 1ß figures directly into Micron's rollout: https://www.micron.com/products/memory/hbm/hbm3e whereas SK Hynix didn't focus on the node but on their heat dissipation technique: https://news.skhynix.com/sk-hynix-develops-worlds-best-performing-hbm3e/ and Samsung similarly while mentioning 1a: https://semiconductor.samsung.com/news-events/tech-blog/leading-memory-innovation-with-hbm3e/.
 
Last edited:
Back
Top