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

[Tuna-Fish]

Why is it only 114.1GB/s for LPDDR6 10.7GT/s 96-bits? It should be 128.4GB/s, which is 50% higher than LPDDR5 10.7GT/s because of the 50% higher width. Is the new algorithm for LPDDR6 resulting in a loss?

LPDDR6 mixes the ECC/DBI bits into the data that's being moved, instead of having separate pins for them. Because of this, it moves 288 bits for each 256 bits of actual data. You have to multiply the width of the interface by 8/9 to get the actual throughput.

This is part of what made non-power-of-2 interface width work. A subchannel moves 12 bits at a time over 24 cycles, for 288 bits total. The size of units CPUs are actually fetching from memory is one cache line, or, on x86, 64B = 512b.

Of that 288, 16 are metadata actually stored on the memory, accessible to the system, and will likely be used for ECC. The other 16 are bits that are generated on transfer by the memory chips, and either used for DBI (reducing power use) or for link protection ECC.

 

[Tigerick]

Why is it only 114.1GB/s for LPDDR6 10.7GT/s 96-bits? It should be 128.4GB/s, which is 50% higher than LPDDR5 10.7GT/s because of the 50% higher width. Is the new algorithm for LPDDR6 resulting in a loss?

Here is official calculator from JEDEC pdf file:

10.667 Gbps x (256 / 288) x (24 / 8) = 28.5 GB/s per 24-bit memory bus

Introductory BW: 12.8 x (256/288) x (24/8) = 34.1 GB/s per 24-bit memory bus (32 GBps)

High End BW: 14.4 x (256/288) x (24/8) = 38.4 GB/s


Thus 96-bit memory bus with 10.667 Gbps LPDDR6 = 28.5 x 4 = 114 GB/s; 12.8 = 136.4 GB/s; 14.4 = 153.6 GB/s.

I have listed all the figures with 96-bit, 192-bit and 384-bit in the table on main page. Have a look!

 

[511]

10.667 Gbps x (256 / 288) x (24 / 8) = 28.5 GB/s per 24-bit memory bus

Thus 96-bit memory bus with 10.667 Gbps LPDDR6 = 28.5 x 4 = 114 GB/s

this would mean a 384bit would be roughly 455GB/s only RIP.
A 256 bit LPDDR5X connection Yields: 256/8*10667 = 341GB/s
A 384 bit LPDDR6 connection Yields: 10667 Gbps x (256 / 288) x (384/ 8) = 455GB/s

 

[511]

maybe we can have this pinned?
@Tigerick

https://www.jedec.org/sites/default/files/Brett%20Murdock_FINAL_Mobile_2024.pdf

 

[Tigerick]

I didn't follow this. Why can't AT3 & AT4 be stand alone GPU dies ?

I think I have answer for your question. First, table times with RDNA5 lineup:

Model	CU	Memory	Memory BW	IC		RDNA5	CU	Memory	Memory BW	IC
Radeon 9070 XT	64	16 GB 256-bit GDDR6	644.6 GB/s	64 MB		AT2- 70XT	64 - 70	18 GB 192-bit GDDR7	864 GB/s	24 MB
Radeon 9070	56	16 GB 256-bit GDDR6	644.6 GB/s	64 MB
Radeon 9070GRE	48	12 GB 192-bit GDDR6	432 GB/s	48 MB		AT2- 70GRE	48	15 GB 160-bit GDDR7	720 GB/s	20 MB
Radeon 9060 XT	32	8/16 GB 128-bit GDDR6	322 GB/s	32 MB		AT2- 60XT	44	12 GB 128-bit GDDR7	576 GB/s	16 MB
Radeon 9060	28	8 GB 128-bit GDDR6	322 GB/s	32 MB		Radeon 9050 ?	32	8/16 GB 128-bit GDDR6	322 GB/s	32 MB
Radeon 7400	28	8 GB 128-bit GDDR6	288 GB/s	32 MB

• See, so simple with the table provided by MLID. AT2 GPU die will be used by Xbox Next and RDNA5 discreet GPU. It helps AMD to spread high cost of N3 process.
• Specs wise: 70CU seems to be max die. AMD most likely use that one as 70XT, thus MLID's claim of 20% faster than RTX4080 seems valid. IC might cut by half, not sure higher speed of GDDR7 will remedy the loss?
• The rumored die size of 264 mm2 might not 100% correct but due to reduction of IC and memory bus, die area should be much smaller than N48 - 357 mm2.
• Above is AT2 die, how about AT3 and AT4? Unlike AT2, AT3 and AT4 should integrate IOD and NPU. If AMD has no other uses, why not just name the die as Medusa Halo? There must have some usage other than APU for Zen 6? Well, I think I am the only one who could answer that......It is going to use on APU for ARM platform. NOT Soundwave, but successor of Soundwave. Yep, AMD won't release just one ARM APU. It is going to be full transition. And that's why I name AMD ARM thread as The Beginning of Transformation !!! IF you don't believe, fine by me. Times will prove my points. And please checkout N1x before comment.
• adroc as usual blindly believe in roadmap without his own judgments. Now I laid my points, adroc what do you think?

 

[branch_suggestion]

• See, so simple with the table provided by MLID. AT2 GPU die will be used by Xbox Next/PS6? and RDNA5 discreet GPU. It helps AMD to spread high cost of N3 process.

Yes, now why wouldn't AMD also try to do the same for AT3/4?

• Specs wise: 70CU seems to be max die. AMD most likely use that one as 70XT, thus MLID's claim of 20% faster than RTX4080 seems valid. IC might cut by half, not sure higher speed of GDDR7 will remedy the loss?

70 is not divisible by 4, it has to be 72. Also MALL is kill outside of CDNA for now, just a fat L2, my guess is 8MB per 64b of G7 and 4MB per 64b of L5X, assuming they use the combo PHY which is 256b L5X/384b L6

• The rumored die size of 264 mm2 might not 100% correct but due to reduction of IC and memory bus, die area should be much smaller than N48 - 357 mm2.

It is also N3P and the uncore is on a seperate die. Seems doable.

• Above is AT2 die, how about AT3 and AT4? Unlike AT2, AT3 and AT4 should integrate IOD and NPU.

NPU is dead, and yes it will have the necessary SoC stuff on die like the Z6LP complex and all of the usual GPU uncore.

• If AMD has no other uses, why not just name the die as Medusa Halo?

Because the other use is being sold as a standalone dGPU?
Not exactly hard to figure out. Yes they lose a bit of margin compared to a pure dGPU part but the reuse on mobile SoCs is well worth it.

• There must have some usage other than APU for Zen 6? Well, I think I am the only one who could answer that......It is going to use on APU for ARM platform. NOT Soundwave, but successor of Soundwave. Yep, AMD won't release just one ARM APU. It is going to be full transition. And that's why I name AMD ARM thread as The Beginning of Transformation !!! IF you don't believe, fine by me. Times will prove my points. And please checkout N1x before comment.

I checked N1X, it is a cost optimised GB10, good GPU, meh CPU for a 2025 part.
Oh wait sorry it got delayed to 2026, not even Mediatek can overcome the Tegra curse.
Give a genuine argument why AMD should go all in on WoA and abandon decades of x86 entrenchment?
November 11 will likely confirm that Zen 7 exists and it is x86, and I get to say I told you so.
Soundwave is a semicustom APU for Microsoft Surface WoA products, that is quite literally it. Van Gogh was Surface semicustom too before presumably Intel did Intel things again and Valve bailed AMD out.
Oh and the last thing, ARM are about to go on the warpath with a full spectrum of first party merchant Si, competing against their own customers since they want to make real money for the first time.
Soon it will not be cheap to use ARM at all, at that point they are little different to the business model of x86.

I thought the $INTC diehards on twitter were unhinged, but they always have good competition.

 

[DavidC1]

Here is official calculator from JEDEC pdf file:

10.667 Gbps x (256 / 288) x (24 / 8) = 28.5 GB/s per 24-bit memory bus

So they are losing a bit on the ECC I'm guessing.

 

[Kepler_L2]

70 is not divisible by 4, it has to be 72. Also MALL is kill outside of CDNA for now, just a fat L2, my guess is 8MB per 64b of G7 and 4MB per 64b of L5X, assuming they use the combo PHY which is 256b L5X/384b L6

I now believe it's actually

8 SE * 2 SA * 6 CU for AT0
4 SE * 2 SA * 5 CU for AT2
2 SE * 2 SA * 6 CU for AT3
1 SE * 2 SA * 6 CU for AT4

With the new gfx13 CU being equivalent to old WGP

L2 should be 4MB per UMC, so 64MB/24MB/32MB/16MB for ATx lineup.

 

[poke01]

Oh wait sorry it got delayed to 2026, not even Mediatek can overcome the Tegra curse.

Mediatek isn’t a good CPU designer, combine them with Tegra folks and you get delays. Xiaomi made a better X925 CPU on the same process.

Give a genuine argument why AMD should go all in on WoA and abandon decades of x86 entrenchment?
November 11 will likely confirm that Zen 7 exists and it is x86, and I get to say I told you so

Zero reason for AMD ditch x86 as they control it.

People need to understand, why Apple moved. Not because ARM is the better ISA, Apple wanted greater control than what AMD and Intel could offer them and moving to in-house ARM designs provided that.

OT Soundwave, but successor of Soundwave. Yep, AMD won't release just one ARM APU. It is going to be full transition. And that's why I name AMD ARM thread as The Beginning of Transformation !!! IF you don't believe, fine by me. Times will prove my points. And please checkout N1x before comment.

 

[adroc_thurston]

Apple wanted greater control

margins.
It's for margins.

 

[poke01]

margins.
It's for margins.

Yeah that too. Margins, control and unification of the code base with the iPhone.

 

[eek2121]

margins.
It's for margins.

This, but yes, also control. Also, the whole x86 vs ARM argument is literally decades old. Give it a rest already. x86 has been dead since the 80s! Long live x86! Meanwhile, AMD just keeps moving forward, taking all the marketshare that everyone claims Qualcomm should be taking right about now.

Maybe just wait for it to finally happen and gloat after? 😉

With specific regards to the topic: We are likely 2-3 years away from widespread DDR6 adoption, and more than a year away from widespread LPDDR6. My own speculation, of course. 😉

 

[johnsonwax]

margins.
It's for margins.

Not here. Cook Doctrine:

We believe that we need to own and control the primary technologies behind the products we make, and participate only in markets where we can make a significant contribution.

Every time Apple has an opportunity to gain an IP advantage, which their ARM architecture license allowed them to hold, they will take it. There aren't meaningful margin opportunities on the silicon, but there is meaningful control opportunities. They build margins off of the control. x86 doesn't allow for this.

It's hard to say what the margin opportunities are in a neutral way, sure they don't have to pay AMDs margins, but there are substantial economies of scale in silicon that going your own way make difficult to match. A series has huge volume, but M series not nearly as much.

 

[branch_suggestion]

I now believe it's actually

8 SE * 2 SA * 6 CU for AT0
4 SE * 2 SA * 5 CU for AT2
2 SE * 2 SA * 6 CU for AT3
1 SE * 2 SA * 6 CU for AT4

With the new gfx13 CU being equivalent to old WGP

L2 should be 4MB per UMC, so 64MB/24MB/32MB/16MB for ATx lineup.

Yay a not completely unhinged lineup. AT3/4 having 1SE would've been weird.
So in effect the CU is now WGP thicc and now CU mode is good enough that WGP mode has no reason to exist?
So AT2 would be 80 old CU equivalent.

 

[Kepler_L2]

Yay a not completely unhinged lineup. AT3/4 having 1SE would've been weird.
So in effect the CU is now WGP thicc and now CU mode is good enough that WGP mode has no reason to exist?
So AT2 would be 80 old CU equivalent.

Yeah, MI400 has the WGP-sized CU with WGP mode and also wave64 support deprecated, and I think this carries over to gfx13.

 

[adroc_thurston]

Yeah, MI400 has the WGP-sized CU with WGP mode and also wave64 support deprecated, and I think this carries over to gfx13.

yeah, these have completely different memory architecture from everything before

 

[johnsonwax]

With specific regards to the topic: We are likely 2-3 years away from widespread DDR6 adoption, and more than a year away from widespread LPDDR6. My own speculation, of course. 😉

I think longer than that, depending on your definition of 'widespread'. So much of the high volume low end PC space really hangs on outdated tech to keep costs down. On Amazon there are still 85 pages of products using LPDDR4, and 22 pages using LPDDR3, compared to 225 pages of LPDDR5 - which is now 5+ years old. iPhone picked up LPDDR5 3 years ago. All rumors are that M5 will pick it up so Apple will have it across their line, and we will see the performant x86 silicon pick it up, but they're gong to have a lot of new silicon that won't use it for quite a while because that's how the SKU spam works - you need to throw off a lot of LPDDR5 SKUs to justify upselling to the LPDDR6 ones.

 

[Doug S]

All rumors are that M5 will pick it up so Apple will have it across their line

Huh? All Apple Silicon uses LPDDR5 or 5X other than the base M1. Or are you saying M5 will use LPDDR6? Not a chance in hell. I would bet heavily against M6/A20P using it either. Apple was slow to adopt LPDDR4/X, and they were slow to adopt LPDDR5/X.

Their model relies on prepaying for massive guaranteed quantities of DRAM, not pushing the envelope and paying more to be an early adopter - and taking the risk of product delays if the promised quantities can't be delivered. They aren't even using the fastest LPDDR5X available, so if they did see a reason to adopt LPDDR6 more quickly it wouldn't be about bandwidth (though it appears early LPDDR6 will be no faster than the fastest LPDDR5X grades, so power savings or ECC would be the only possible motivation)

 

[johnsonwax]

Huh? All Apple Silicon uses LPDDR5 or 5X other than the base M1. Or are you saying M5 will use LPDDR6? Not a chance in hell. I would bet heavily against M6/A20P using it either. Apple was slow to adopt LPDDR4/X, and they were slow to adopt LPDDR5/X.

Their model relies on prepaying for massive guaranteed quantities of DRAM, not pushing the envelope and paying more to be an early adopter - and taking the risk of product delays if the promised quantities can't be delivered. They aren't even using the fastest LPDDR5X available, so if they did see a reason to adopt LPDDR6 more quickly it wouldn't be about bandwidth (though it appears early LPDDR6 will be no faster than the fastest LPDDR5X grades, so power savings or ECC would be the only possible motivation)

Rumors are that M5 will use LPDDR6 and A series next year. I can't assess the credibility of those rumors, just that I've seen them repeatedly.

Apple does sometimes skip generations when there are architectural benefits to doing that, they also do push the envelope to be an early adopter when they can dominate the supply side of that market (we would know if they had done that here). A7 was an early adopter jump to 64bit ARM so that they could start the process of transitioning developers off of 32bit with the expectation that they'd remove them entirely from their platform, and so they could drive consumer expectations for a few cycles. Apple was the first to ship NPUs in volume because they had a specific need. iPhones adopted NVMe sooner than most PCs did for similar reasons. I'll note that the place where Apple Silicon has diverged most aggressively from x86 is in the memory model and memory access, and this is the first memory architecture change that has arrived on Apple's M series horizon. I'm just saying that there is no past history to inform us here. This is really the first data point to inform us with respect to M series. What we also do know about Apple is that they pick their battles very carefully. They tend to be conservative with marginal benefits and risk favored with non-marginal ones. They're not going to waste engineer time on small gains when there are big ones on the horizon - they have always run quite lean on engineering staffing for a company their size, and that's just a consequence of that - you can't jump at everything - that's Steve Jobs' focus quote right there - that 5X launched right after M series likely meant that Apple just didn't have the resources to do it all.

Note, Apple doesn't have benefit of access to DDR6 on GPUs for AI like the PC market has, and has had for several years. Toward Apple's panic on AI (which I think is unfounded, but they don't) getting LPDDR6 may be seen as quite important because they have to do it on the back of their unified memory model and because of their commitment to doing it on-device. There's a fairly persistent and specific rumor that Samsung is working on LPDDR6-PIM for Apple to ship as early as next year, and we already have reports of increased prepayments to Samsung for memory for this year's cycle (Samsung is providing more of this years memory at the expense of SK Hynix/Micron) which could be hiding such an investment in LPDDR6 in the next cycle. (I've been caught out by these in the past - where a step change in prepayments came earlier than the technological cycle and caused people to think that they weren't related, when they were simply because that was part of a broad multi-year agreement.) Anyway, I think there are reasons why today Apple would be more concerned with being in a leadership position on memory, there's a prepayment step change with Samsung that could (or not) correlate to such a move. So there's motive and opportunity.

Not saying these are credible rumors, but I don't see arguments to discredit them.

 

[Tigerick]

Rumors are that M5 will use LPDDR6 and A series next year. I can't assess the credibility of those rumors, just that I've seen them repeatedly.

I thought I was the only one who is expecting M5 series to support LPDDR6. Any speculation on what big changes on M5 series?

Apple does sometimes skip generations when there are architectural benefits to doing that, they also do push the envelope to be an early adopter when they can dominate the supply side of that market (we would know if they had done that here). A7 was an early adopter jump to 64bit ARM so that they could start the process of transitioning developers off of 32bit with the expectation that they'd remove them entirely from their platform, and so they could drive consumer expectations for a few cycles. Apple was the first to ship NPUs in volume because they had a specific need. iPhones adopted NVMe sooner than most PCs did for similar reasons. I'll note that the place where Apple Silicon has diverged most aggressively from x86 is in the memory model and memory access, and this is the first memory architecture change that has arrived on Apple's M series horizon. I'm just saying that there is no past history to inform us here. This is really the first data point to inform us with respect to M series. What we also do know about Apple is that they pick their battles very carefully. They tend to be conservative with marginal benefits and risk favored with non-marginal ones. They're not going to waste engineer time on small gains when there are big ones on the horizon - they have always run quite lean on engineering staffing for a company their size, and that's just a consequence of that - you can't jump at everything - that's Steve Jobs' focus quote right there - that 5X launched right after M series likely meant that Apple just didn't have the resources to do it all.

Note, Apple doesn't have benefit of access to DDR6 on GPUs for AI like the PC market has, and has had for several years. Toward Apple's panic on AI (which I think is unfounded, but they don't) getting LPDDR6 may be seen as quite important because they have to do it on the back of their unified memory model and because of their commitment to doing it on-device. There's a fairly persistent and specific rumor that Samsung is working on LPDDR6-PIM for Apple to ship as early as next year, and we already have reports of increased prepayments to Samsung for memory for this year's cycle (Samsung is providing more of this years memory at the expense of SK Hynix/Micron) which could be hiding such an investment in LPDDR6 in the next cycle. (I've been caught out by these in the past - where a step change in prepayments came earlier than the technological cycle and caused people to think that they weren't related, when they were simply because that was part of a broad multi-year agreement.) Anyway, I think there are reasons why today Apple would be more concerned with being in a leadership position on memory, there's a prepayment step change with Samsung that could (or not) correlate to such a move. So there's motive and opportunity.

Not saying these are credible rumors, but I don't see arguments to discredit them.

These, Apple M1 Pro and Max are the first to support up to 512-bit LPDDR5-6400 and until today, there is no player catching up yet. The next bar would be supporting 384-bit LPDDR6. However, competition landscape will change by the end of this year with ARMies incoming.

 

[Tigerick]

they were slow to adopt LPDDR5/X.

Apple's M1 Pro and Max are the first to adopt LPDDR5, please check the fact before posting. I have posted LPDDR5 launching timeline on the main page, have a look.

As for LPDDR5x supporting M4 series, please check the difference between M3 Max and M4 Max, hope you get the reason behind.

 

[511]

I thought I was the only one who is expecting M5 series to support LPDDR6. Any speculation on what big changes on M5 series?

N3E -> N3P expect 5-10% performance improvements

 

[Doug S]

I thought I was the only one who is expecting M5 series to support LPDDR6. Any speculation on what big changes on M5 series?

When I saw @johnsonwax mention he'd seen that rumor I had to think about how I recently saw a story stating that DDR6 will use 24 bit channels. That's something I said here and seen no one else suggest, and either it was picked up by someone who read it or (perhaps more likely) was scraped by some AI. The same probably happened with what you said about M5 and LPDDR6.

For a "dead" forum, we sure seem to be have a lot of control over the rumor narrative! At least when it comes to DRAM.

I'm gonna have to give some thought to how we could start a rumor here that gets picked up in the same way, that we can use for some advantageous stock trading. Can't charge us with insider trading for speculating, and some AI scrapes our comments, then some journalist doing "research" asks an AI a question and our rumors get repeated as the truth!

 

[johnsonwax]

When I saw @johnsonwax mention he'd seen that rumor I had to think about how I recently saw a story stating that DDR6 will use 24 bit channels. That's something I said here and seen no one else suggest, and either it was picked up by someone who read it or (perhaps more likely) was scraped by some AI. The same probably happened with what you said about M5 and LPDDR6.

For a "dead" forum, we sure seem to be have a lot of control over the rumor narrative! At least when it comes to DRAM.

I'm gonna have to give some thought to how we could start a rumor here that gets picked up in the same way, that we can use for some advantageous stock trading. Can't charge us with insider trading for speculating, and some AI scrapes our comments, then some journalist doing "research" asks an AI a question and our rumors get repeated as the truth!

So, the rumor I've heard is 96bit/192/384 for the M5/Pro/Max. That's upwards of a year old so I don't know how it interacts with the timing of everyone else's stuff. It may or may not be the source for all of it, and also be wrong. The recent story of 24bit channels, if correct and independent, confirms that rumor to some degree. Because Apple is designing their own memory controllers, I think it makes sense (just from an ROI perspective) for them to jump on these things early and ride them long. I didn't make much of this particular rumor because it seemed pretty obvious. I find the Samsung PIM rumor to be a lot more interesting, and that rumor specifically mentions its targeting iPhone which would put iPhone on LPDDR6 as early as next year. That makes sense to me because the iPhone was the TPU driver for Apple due to the computational photography task and the need for that to be extremely responsive due to often doing it in-viewfinder, so it needs to be realtime. So wanting to allow that computational work to happen with a larger sensor (which needs geometrically more compute power) would push Apple to put LPDDR6 on the iPhone sooner than you might predict because shoving ANE into RAM probably solves that problem, and if you need to design a memory controller for it, why not go the full monty, do your whole lineup, and put all of that in your prepayment contract with Samsung? Even if it's only marginally beneficial for M series, it'd be hugely beneficial for A series for that application, and that's why you do it ASAP. iPhone got NVMe sooner than most PCs because getting data off the camera sensor into storage was the problem to be solved and NVRAM at the time wasn't going to be able to eat 60FPS off of 3 high resolution cameras. Your budget PC never needs to do that. So the iPhone camera has been one of Apple's main technological drivers for some time now and that Samsung rumor fits right into that. Everything else works backward from that.

 

[Doug S]

So, the rumor I've heard is 96bit/192/384 for the M5/Pro/Max. That's upwards of a year old so I don't know how it interacts with the timing of everyone else's stuff. It may or may not be the source for all of it, and also be wrong. The recent story of 24bit channels, if correct and independent, confirms that rumor to some degree.

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.