Info LPDDR6 @ Q3-2025: Mother of All CPU Upgrades

Page 12 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

johnsonwax

Senior member
Jun 27, 2024
403
604
96
Not really. LPDDR5/X controllers are 16 bit, so you can do 96/192/384 with those just as easily as you can with 24 bit. In fact Apple already shipped Apple Silicon SoCs with 192 bit wide memory bus - using LPDDR5X.
Except the rumor was explicitly mentioning LPDDR6. So the source of the rumor correctly guessed that those were viable memory bus sizes. Maybe that's obvious from LPDDR5X but it's also what people would have assumed from 24 bit channels. That's what was guessed upthread before I brought in the roughly year-old rumor that said the same thing.

And that still leaves us with the separate Samsung LPDDR6-PIM rumor which is directed at next year's iPhones. You think iPhone might get LPDDR6 before M series? I don't.
 

Doug S

Diamond Member
Feb 8, 2020
3,617
6,393
136
Except the rumor was explicitly mentioning LPDDR6. So the source of the rumor correctly guessed that those were viable memory bus sizes. Maybe that's obvious from LPDDR5X but it's also what people would have assumed from 24 bit channels. That's what was guessed upthread before I brought in the roughly year-old rumor that said the same thing.

And that still leaves us with the separate Samsung LPDDR6-PIM rumor which is directed at next year's iPhones. You think iPhone might get LPDDR6 before M series? I don't.

This might be another case where the rumors come from posts made here. I believe @Tigerick speculated well over a year ago about iPhones using 96 bits of LPDDR6. AI scrapes posts here, then his speculation about this becomes fact when someone asks AI a question, just like I saw my speculation about DDR6 using 24 bit subchannels posted as "fact" somewhere recently despite the DDR6 spec not being out yet.

I'm doubly skeptical of claims that Apple is using PIM technology - and even if so certainly not as soon as next year.
 

johnsonwax

Senior member
Jun 27, 2024
403
604
96
This might be another case where the rumors come from posts made here. I believe @Tigerick speculated well over a year ago about iPhones using 96 bits of LPDDR6. AI scrapes posts here, then his speculation about this becomes fact when someone asks AI a question, just like I saw my speculation about DDR6 using 24 bit subchannels posted as "fact" somewhere recently despite the DDR6 spec not being out yet.

I'm doubly skeptical of claims that Apple is using PIM technology - and even if so certainly not as soon as next year.
Could be. Though iPhone is an ideal candidate for PIM because it's so heavily built around computational photography and bandwidth issues from the camera sensor have been a primary driver of the platform, and it's got enough volume to justify a dedicated solution. It's not like RAM is a consumer variable in the product either. It usually takes us weeks to even learn how much RAM is on board.

I suspect you would have been skeptical of Apple shipping an NPU first on iPhone in 2018, and yet here we are.
 

Doug S

Diamond Member
Feb 8, 2020
3,617
6,393
136
Could be. Though iPhone is an ideal candidate for PIM because it's so heavily built around computational photography and bandwidth issues from the camera sensor have been a primary driver of the platform, and it's got enough volume to justify a dedicated solution. It's not like RAM is a consumer variable in the product either. It usually takes us weeks to even learn how much RAM is on board.

I suspect you would have been skeptical of Apple shipping an NPU first on iPhone in 2018, and yet here we are.

It is very very difficult to do logic on a DRAM process. They'd have to make major changes to the process, but that would massively compromise the density achievable with DRAM because you couldn't do trenching. At that point you might as well do your PIM with SRAM.

Maybe the density of TSVs is sufficient now that a logic chip could become part of a DRAM stack, but it would still compromise DRAM density having to make room for hundreds of thousands of TSVs - and those would be non standard parts that by definition cost a lot more than mass manufactured DRAM.
 

johnsonwax

Senior member
Jun 27, 2024
403
604
96
It is very very difficult to do logic on a DRAM process. They'd have to make major changes to the process, but that would massively compromise the density achievable with DRAM because you couldn't do trenching. At that point you might as well do your PIM with SRAM.

Maybe the density of TSVs is sufficient now that a logic chip could become part of a DRAM stack, but it would still compromise DRAM density having to make room for hundreds of thousands of TSVs - and those would be non standard parts that by definition cost a lot more than mass manufactured DRAM.
Yeah, I think it's conceivable because it's iPhone volume, and it's Apple. I don't see it making sense for the Mac. But Apple knows the big cycle driver for iPhone is the camera and so if they have some computational photography options that PIM could unlock, that's 100 million Pro units a year. iPhone is its own economy of scale. But it only makes sense to me if it unlocks something new. It needs to be an enabling tech to make sense, not a marginal one. That's why Apple put ANE in the A series to begin with, etc. Their tentpole feature didn't work without it.
 

marees

Golden Member
Apr 28, 2024
1,804
2,415
96
Why is AMD using LPDDR for dGPUs now...and not before? (Timestamped video)

 
Last edited:

NTMBK

Lifer
Nov 14, 2011
10,451
5,836
136
This might be another case where the rumors come from posts made here. I believe @Tigerick speculated well over a year ago about iPhones using 96 bits of LPDDR6. AI scrapes posts here, then his speculation about this becomes fact when someone asks AI a question, just like I saw my speculation about DDR6 using 24 bit subchannels posted as "fact" somewhere recently despite the DDR6 spec not being out yet.

I'm doubly skeptical of claims that Apple is using PIM technology - and even if so certainly not as soon as next year.
I'm pretty sure this has happened a few times now. If we start seeing articles speculating about FD-SOI and CMT then we'll know for sure
 

Doug S

Diamond Member
Feb 8, 2020
3,617
6,393
136
I'm pretty sure this has happened a few times now. If we start seeing articles speculating about FD-SOI and CMT then we'll know for sure

It probably has, but I think AI makes that sort of thing much more likely. At heart an LLM is a text prediction machine, it doesn't understand what it is ingesting thus can't easily tell the difference between a post reporting a fact and a post making an educated guess (or a wild guess or deliberate lie, for that matter) A human reading my posts regarding DDR6 will understand that I'm not speaking from a position of knowledge - the DDR6 specs haven't been released and I have never claimed to be on the standards committee or know anyone who is so there is no reason I would know its specs for a fact. So at most they'd write an article saying "some people are speculating that ..." not reporting it as fact.

If the AI isn't ingesting a lot of content where the token sequences for "DDR6" and "subchannels" in the context of bit width occur in close proximity (which it very likely is not) it will only take a few posts from me doing so to get the AI to predict "24 bit subchannels" follows "DDR6" in its responses. For something with a lot more token sequences out there like "Zen 6" me posting "Zen 6 will be 70% faster than Zen 5" isn't going to affect the AI's weightings. But if that became a joke so everyone was posting it not just here but on other forums to where it overwhelmed all other sequences suggesting Zen 6's performance improvement, you might get that answer if you asked an AI how much faster Zen 6 is expected to be.
 
  • Like
Reactions: 511

Tigerick

Senior member
Apr 1, 2022
859
804
106
LPDDR6-HVM.jpg
According to Bilibili, Samsung has finally revealed their LPDDR6 HVM date: end of 2025. The issue is Samsung only could HVM lowest speed bin, that would be 10.7Gbps. Judging by the low yield of HBM4, we shouldn't be surprised by that. Let's hope Hynix and Micron are able to provide high speed bin of LPDDR6 next year...
 

Doug S

Diamond Member
Feb 8, 2020
3,617
6,393
136
LPDDR6 rollout is gonna be slow. All the DRAM vendors are concentrating on HBM because that's where the gold mine is.
 

Doug S

Diamond Member
Feb 8, 2020
3,617
6,393
136
What? DDR6 on AMD is only 2030?

JEDEC hasn't even published DDR6, and the time between LPDDR6 being published and its final approval was ~1 year. Synopsys just released the first PHY, Samsung just announced the start of mass production. So realistically we aren't going to see anything that uses it until this time next year at the earliest - and the first version is rather meh since it is only -10666 and you can already get LPDDR5X/5T and that speed and higher. So I'm not holding my breath for anything other than maybe some niche China phones looking for something to grab attention/press using it next year.

So if they publish the DDR6 spec this year and the final approval is by the end of next year, the earliest systems using it would appear is early 2028. But probably not that soon since it will (just like DDR5, DDR4 etc. before it) appear in servers first and they have longer qualification cycles. So this time in 2028 for servers is more realistic. So maybe sometime in 2029 for desktop would be a good guess, though that might be for Threadripper type systems that can really benefit from the extra bandwidth and mainstream desktops take longer.
 

regen1

Member
Aug 28, 2025
150
219
71
and the first version is rather meh since it is only -10666 and you can already get LPDDR5X/5T and that speed and higher.
You are still getting a third more B/W at the same frequencies.
For eg. standard dual channel LPDDR5X @10667MT/s will get ~170GB/s,
standard dual channel LPDDR6 @10667MT/s will get ~227.5GB/s.

But yeah LPDDR6 is still quite some time away for being available and then become mainstream. DDR6 seems even further away for availability .
 

511

Diamond Member
Jul 12, 2024
4,641
4,241
106
I think after some time we should really change the platform 5-6 years should be enough for a platform.