Friday, November 7th 2025

Samsung Prepares 10.7 Gbps LPDDR6 Memory Built on 12 nm Node

Samsung has developed its next-generation low-power LPDDR6 memory for mobile applications. Built on a 12 nm process, this LPDDR6 memory module achieves data rates of 10.7 Gbps. Engineers have increased the input/output channel count and introduced dynamic power tuning, which adjusts power consumption in real time based on workload demands. Samsung has also incorporated robust security features directly into the memory subsystem to ensure data integrity across all applications. Interestingly, Samsung's improved and tuned LPDDR5X also reaches 10.7 Gbps, which is the starting point for the new LPDDR6 memory.

As seen in the transition from DDR4 to DDR5, the final configurations of the previous generation memory are often close in performance to the first generation of the new standard. For LPDDR6, Samsung claims a 21% improvement in energy efficiency over its predecessor, LPDDR5X. This means devices can achieve the same bandwidth and speed while using one-fifth less power. Future LPDDR6 iterations are expected to reach speeds of up to 14 Gbps, surpassing even the fastest overclocked LPDDR5X chips. More details are anticipated at this year's CES, where Samsung will showcase devices powered by the new memory and reveal specific features.
Sources: CES, @harukaze5719
Add your own comment

14 Comments on Samsung Prepares 10.7 Gbps LPDDR6 Memory Built on 12 nm Node

#1
TechBuyingHavoc
Interesting that this memory is going to be produced on a inferior node (12nm) vs LPDDR5 (10nm nodes). Capacity issues?
Posted on Reply
#2
Darkmont
TechBuyingHavocInteresting that this memory is going to be produced on a inferior node (12nm) vs LPDDR5 (10nm nodes). Capacity issues?
That is their best node, when memory manufacturers say "10nm" it's in reference to a class of node (I.E. 17-18nm, 15-16, 12-14nm, etc.).
Posted on Reply
#3
TechBuyingHavoc
DarkmontThat is their best node, when memory manufacturers say "10nm" it's in reference to a class of node (I.E. 17-18nm, 15-16, 12-14nm, etc.).
Hmm, that is good to know, I was wondering why the node would go backwards. So a 20nm class node would be everything from 20nm to 29nm?
Posted on Reply
#4
bonehead123
Gimme (moar) suppeeeeed, or gimme death, hahahaha :D

Just do a tush push on the ram chips & be happy !..
Posted on Reply
#5
shuffy_losys
TechBuyingHavocInteresting that this memory is going to be produced on a inferior node (12nm) vs LPDDR5 (10nm nodes). Capacity issues?
The main problem is DRAM doesn't scale very well below 14nm class process nodes due to the capacitor element.
Posted on Reply
#6
DavidC1
shuffy_losysThe main problem is DRAM doesn't scale very well below 14nm class process nodes due to the capacitor element.
12nm DRAM is way smaller than logic processes. That's why they hit the scaling wall long time ago. Logic is still scaling because they aren't at that level of density yet. Look at Intel's eDRAM. It's on processes optimized for logic.
www.realworldtech.com/intel-dram/2/
17.5Mb/mm2 on 22nm.
semiwiki.com/forum/threads/techinsights-samsung-20-nm-ddr4-tsv-enabled-dram.5783/
1GB(8Gb) on 56mm2 die versus 128MB on 80mm2 die. It's 10x the density. This means a N3 logic node is needed to be at equal level density to 10nm DRAM node.
Posted on Reply
#7
shuffy_losys
DavidC112nm DRAM is way smaller than logic processes. That's why they hit the scaling wall long time ago. Logic is still scaling because they aren't at that level of density yet. Look at Intel's eDRAM. It's on processes optimized for logic.
The problem is that for there to be a meaningful amount of charge in the capacitor, it must be at least a certain volume. If you shrink the lateral size of the capacitor, you have to increase its depth to retain the meaningful amount of charge. See blocksandfiles.com/2020/04/13/dram-is-stuck-in-a-10nm-process-trap/ and www.appliedmaterials.com/us/en/blog/blog-posts/dram-scaling-requires-new-materials-engineering-solutions.html
Posted on Reply
#8
DavidC1
shuffy_losysThe problem is that for there to be a meaningful amount of charge in the capacitor, it must be at least a certain volume. If you shrink the lateral size of the capacitor, you have to increase its depth to retain the meaningful amount of charge. See blocksandfiles.com/2020/04/13/dram-is-stuck-in-a-10nm-process-trap/ and www.appliedmaterials.com/us/en/blog/blog-posts/dram-scaling-requires-new-materials-engineering-solutions.html
So why is logic gains also crashing at N2/18A? N2 gains almost nothing in density while 18A is 30%, which includes PowerVia gains. That's not a new node. That's not even a half node gain. It used to be 100% density gain. Things more fundamental than a capacitor is the limit here. It's like there's a finish line and DRAM had only 10m to go while logic had 30m more to go.
Posted on Reply
#9
Wirko
Static RAM also stopped scaling down, or almost, a few nodes ago. I made a rough calculation and comparison about two years ago, using the scarce available data about densest nodes in production. One SRAM bit (6 transistors) takes up the same space as ~12 DRAM transistors. I don't understand the structure of DRAM transistor well enough but it seems to be vertical, therefore taking up less area than a SRAM/logic transistor would.
Both SRAM and DRAM are now made with EUV and both hardly scale anymore. To me it means that DRAM cell shrinking is limited both by minimum capacitor size and fab resolution.

Still, 4F²-sized DRAM cells are being researched, which would shrink the cell size by a third compared to today's 6F². Researchers know that the capacitor's area would have to be reduced by the same amount, so they probably have some ideas how to do that.
Posted on Reply
#10
TechBuyingHavoc
DavidC1So why is logic gains also crashing at N2/18A? N2 gains almost nothing in density while 18A is 30%, which includes PowerVia gains. That's not a new node. That's not even a half node gain. It used to be 100% density gain. Things more fundamental than a capacitor is the limit here. It's like there's a finish line and DRAM had only 10m to go while logic had 30m more to go.
Well, there are some economical limits and two intrinsic physical limits. The first physical limit is that electrons start tunneling through transistor structures in ways that are not intended (like water leaking through dam walls) as the structures continue to shrink over time. This requires special materials (high-K layers), different design rules, more consideration on a thousand different variables that previously was less of an issue. All of that adds cost and time which affects the economical limit.

The 2nd physical limit is the wavelength of light which is several times larger than the size of the structures that this light is used to construct through lithography. We now have EUV and perhaps soft X-Ray machines in the future that shrink the wavelength of light but at some point, this will not shrink any further and we will be stuck permanently. There are workarounds for now like double-patterning and quad-patterning that delay the problem but it is still there and the workarounds add cost and time (which adds to the cost again).

A earlier form of scaling, Dennard scaling where the power usage of the transistor falls as the transistor size shrinks, already died in 2003. That started to increase costs as chip designers had to come up with power gating, dark silicon, etc to deal with a constrained power budget (which adds both development and product costs again).

This all feeds to the economical limit which is the cost of developing a new node, or a new product on that node. Both fab development costs and product development costs are increasing over time (related to the physical limit). This used to be compensated by shrinking the chip area, allowing for more products per wafer. Costs were spread out without cutting production because there were always more people to sell chips to.

There are two further limits on that, one market-related and one chip-design-related. There are not enough people to continue growing product volume so now the profit calculation around new nodes is more complicated. At each new node, customers have been delaying or just straight abandoning the transition to a new node because the economics don't make sense anymore. Only the largest companies have the product sales volumes to justify new nodes (Apple, etc).

The chip-design reason is that most cores spend their time waiting for data to arrive/leave instead of processing that data. Chip designs therefore must add different stages of memory/storage to reduce the time for data travel like cache, RAM, etc. The lowest latency solution is to put the memory directly on the die, right next to the logic transistors waiting for that data. But as others have explained, capacitance relies on internal volume so this shrinks slower than the logic transistors. Over time, more and more of the die area is cache which sabotages the economic scaling for a new node (adding costs again).

TLDR = Making smaller and smaller structures is damn hard and getting harder. All of this difficulty requires money to overcome.

This is all off the top amateur interest knowledge from a non-professional so take it with a grain of salt, I am sure I got half of the details wrong.

EDIT #1 = I feel like I used "limit" incorrectly here, most of these are not hard limits but obstacles to overcome that keep increasing in difficulty.
EDIT #2 = Wire resistance is another physical limit, wire scaling is vastly slower than transistor logic (thank you, Wirko)
Posted on Reply
#11
Wirko
TechBuyingHavocThe first physical limit is that electrons start tunneling through transistor structures in ways that are not intended
There's one more limit: resistance of wires carrying power (but also signals) across the chip. Backside power delivery will help somewhat but it's not a magical solution, and it will come at a high cost.
Posted on Reply
#12
TechBuyingHavoc
WirkoThere's one more limit: resistance of wires carrying power (but also signals) across the chip. Backside power delivery will help somewhat but it's not a magical solution, and it will come at a high cost.
I knew I missed something.
Posted on Reply
#13
Wirko
DarkmontThat is their best node, when memory manufacturers say "10nm" it's in reference to a class of node (I.E. 17-18nm, 15-16, 12-14nm, etc.).
10nm-class is a very broad term, manufacturers usually use 1x, 1y, 1z, 1a, 1b, 1c etc. to refer to many, many integers between 19 and 10. Samsung is the only one that uses actual numbers I think, but not always.

Anyways, has anyone announced any processors with a LPDDR6 memory controller already?
Posted on Reply
#14
megaclite
Samsung not even able to release DDR5-6400 JEDEC
Posted on Reply
Add your own comment
Nov 21st, 2026 12:04 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts