Info LPDDR6 @ Q3-2024: Mother of All CPU Upgrades

[FlameTail]

• X Elite with 8 cores might not be correct naming...the thing is the launching date is same as 8 Gen4, could X without Elite be using the same SoC as 8 Gen4??? And 8 cores SoC could be 2+6 Phoenix cores?

Have you heard of the Snapdragon X Plus?

 

[Tigerick]

Another benefit of 384-bit memory bus compared to 512-bit memory bus is reduction of SLC cache as shown below:-

View attachment 95457
SLC cache is actually associated with memory controller. With 384-bit memory bus, the amount of SLC should drop from 64MB to 48MB.

M3 Max Chip Has Hidden Change That Could Affect Future 'M3 Ultra' Chip

Apple's M3 Ultra chip may be designed as its own, standalone chip, rather than be made up of two M3 Max dies, according to a plausible new theory. The theory comes from Max Tech's Vadim Yuryev, who outlined his thinking in a post on X earlier today. Citing a post from @techanalye1 which...

forums.macrumors.com

92 billions in M3 Max is pretty large number of transistors. In fact, TSMC's roadmap shown 100 billions transistors in monolithics die only available with upcoming N2 process. Apple is pushing the design limit with M3 Max. That's why I suspect Apple might remove ultra fusion connectors in M3 Max cause Apple might skip M3 Ultra in M3 generation.

If Apple is indeed moving M4 series with LPDDR6 in Q1 2025, the changes would be pretty big. That warrants Apple to skip M3 Ultra and moving to M4 Ultra ASAP...




N3B still remain the only process that provides density Apple wants. As I calculated, if Apple is moving to 384-bit LPDDR6 memory bus, Apple could save some die area as shown above...

Of course, this is my speculation: if Apple is not announcing any M3 Ultra products in upcoming WWDC, yeah we should be expecting M4 Ultra in early 2025...

 

[FlameTail]

Apple should max out the reticle limit and make a monolithic 800+ mm² M Ultra chip like Nvidia does with their GPUs.

Edit: or we could go even crazier and make a 1600 mm² monolithic M Extreme chip using reticle stitching.

 

[Tigerick]

Apple should max out the reticle limit and make a monolithic 800+ mm² M Ultra chip like Nvidia does with their GPUs.

Edit: or we could go even crazier and make a 1600 mm² monolithic M Extreme chip using reticle stitching.

And consume at least 700W power TDP, now you know why your response sounds so stupid....

 

[FlameTail]

M3 Max:
CPU: 56W
GPU: 60W
Combination: 116W.

Let's round it upto 125W so that the NPU and other components are also covered.

If the hypothetical M3 Extreme chip is 4× that of the M3 Max, then it's power consumption would also be 4× of M3 Max, which would be 500W.

 

[Doug S]

The Ultra is too low volume for it to make sense to design as a standalone, let alone a reticle sized standalone. The 2xMax solution works well for them because it leverages a higher volume die to make that low volume product. It would be doubly true if they want to go bigger than Ultra (if they don't with M4 I think it is safe to say they have no plans to ever do so)

The reason there are no I/O pads on the bottom of M3 Max are, IMHO, because there will be no M3 Ultra. M2 Ultra came out in June 2023, the M4 family will ship this fall one year after the M3 family, enabling an M4 Ultra before the end of the year. There was no point to slotting an M3 Ultra in that 16-18 month window.

 

[okoroezenwa]

The reason there are no I/O pads on the bottom of M3 Max are, IMHO, because there will be no M3 Ultra. M2 Ultra came out in June 2023, the M4 family will ship this fall one year after the M3 family, enabling an M4 Ultra before the end of the year. There was no point to slotting an M3 Ultra in that 16-18 month window.

You think they’ll skip the M3 Ultra and go with M4 Ultra by the end of this year? I’d be very surprised with that move. Or did I interpret this wrong?

 

[FlameTail]

You think they’ll skip the M3 Ultra and go with M4 Ultra by the end of this year? I’d be very surprised with that move. Or did I interpret this wrong?

Honestly, I could see that happening.

 

[Tigerick]

All premium FF AMD SoCs are 128b.
Don't.

64-bit memory bus will keep improving with each speed upgrade. As you can see with upcoming Sonoma Valley which comes with 64-bit LPDDR5x support, the performance of APU should be comparable to Apple M1 with similar memory bandwidth.

The best has yet to come, with 64-bit LPDDR6, all upcoming APU should offer similar performance of Apple M3. That includes ARM version SoC, upcoming Steam Deck 2 and rumored XBOX Handheld. It seems all SoCs are coming in 2026....can't wait to fill in the blank in the future

64-bit Memory Bus	Memory BW	Model	Date	Node	CPU	GB6 1T	PPC	GB6 MT	GPU	FP32
LPDDR5-5500	44 GB/s	Mendocino	Q3 2022	N6	Zen 2 - 4C	1364	316	4485	RDNA2 - 2 CU	0.49 TF
LPDDR5x-8533	68 GB/s	Sonoma Valley	Q1 2026 ?	SF4X	Zen 5c - 4C				RDNA3+ 4CU ?	~ 2 TF ?
128-bit LPDDR4x-4266	68 GB/s	M1	Q4 2020	N5	4P+4E	2354	736	8404	8 core	2.6 TF
LPDDR6-12800	102 GB/s	ARM	2026 ?	?	8C ?				RDNA 8CU ?	~4 TF ?
		Steam Deck 2	2026 ?	?	?				?	~4 TF ?
		XBOX Handheld	2026 ?	?	?				?	~4 TF ?
128-bit LPDDR5-6400	102 GB/s	M3	Q4 2023	N3B	4P+4E	3084	760	11564	10 core	4.3 TF
96-bit LPDDR6-12800	150 GB/s	M4	Q1 2025 ?	N3B

 

[adroc_thurston]

64-bit memory bus will keep improving with each speed upgrade

This isn't a bandwidth issue.

 

[Tuna-Fish]

LPDDR is strictly point-to-point, and after a point stacking more dies in a package gets more expensive than just having a wider bus. I don't think it would be cost-effective to have 64b LPDDR6 bus for APUs.

 

[Tigerick]

LPDDR is strictly point-to-point, and after a point stacking more dies in a package gets more expensive than just having a wider bus. I don't think it would be cost-effective to have 64b LPDDR6 bus for APUs.

You should at least understand the basic of Steam Deck OLED's improvement which reduce the memory chips from 4 to 2. With 64-bit memory bus, 12GB only need four dies in single chip, thus saving PCB and cost.

 

[okoroezenwa]

Honestly, I could see that happening.

I can see no M3 Ultra but M4 Ultra end of this year in addition would be surprising.

 

[Tigerick]

As @Doug S want to bet against me. I think why not make a table listing all speculated iPhone's RAM configuration. It seems LPDDR6 might be coming by end of this year, that's why I am speculating this year's iPhone Pro 16 will support 12GB LPDDR6. Yeah, big bet. I am betting Apple will skip LPDDR5x/T no matter what, thus either LPDDR5 OR LPDDR6. @Doug S You can use the same table format to list down your RAM configuration and we will see the results by end of this year.

	iPhone	Node	RAM		iPhone Pro	Node	RAM
2023	A16	N5P	6GB LPDDR5		A17 Pro	N3B	8GB LPDDR5
2024	A18	N3E	8GB LPDDR5		A18 Pro	N3E	12GB LPDDR6
2025	A19	N3E	8GB LPDDR6		A19 Pro	N3P	12GB LPDDR6
2026	A20	N3P	12GB LPDDR6		A20 Pro	N2	16GB LPDDR6

 

[FlameTail]

Now make a table detailing the past 5 years of Galaxy S Ultras and iPhone Pro Maxes.

You will see that thr iPhone was always late in adopting a new LPDDR version.

 

[Doug S]

Now make a table detailing the past 5 years of Galaxy S Ultras and iPhone Pro Maxes.

You will see that thr iPhone was always late in adopting a new LPDDR version.

Yep, Apple is always a couple years behind. They have to be in some sense because new LPDDR chips aren't available in enough guaranteed volume in year one for them to include in such a high volume product. But they usually wait two years before pulling the trigger, because it is still more costly and the gains just aren't that great in a phone despite the massive LPDDR6 hype we are getting in this forum.

I think Apple adopts LPDRR5X this fall, which will be about two years after it became available and appeared in the first Androids. They will first adopt LPDDR6 in 2027, about two years after it becomes available and appears in the first Androids.

Android OEMs aren't doing it because the gains are so great, BTW. They are doing it because they know the other guy will be doing it, and Android buyers are much more obsessed with specs due to the number of brands they have to choose from.

 

[FlameTail]

I think Apple adopts LPDRR5X this fall, which will be about two years after it became available and appeared in the first Androids. They will first adopt LPDDR6 in 2027, about two years after it becomes available and appears in the first Androids.

Yeah, I reckon M4 and M5 will be LPDDR5X (or maybe LPDDR5X/5T-9600? Probably not). M6 might be LPDDR6, or LPDDR5X. M7 will quite certainly be LPDDR6.
This is assuming that henceforth, a new M generation comes every 12 months.

Android OEMs aren't doing it because the gains are so great, BTW. They are doing it because they know the other guy will be doing it, and Android buyers are much more obsessed with specs due to the number of brands they have to choose from.

Certainly. And also Apple SoCs huge caches offer sufficient internal bandwidth to offset the external RAM bandwidth disadvantage vs Android SoCs.

 

[Tigerick]

As expected, two persons are using past history try to speculate Apple will be late to adopt LPDDR6. And they are ignoring that how aggressive Apple is adopting LPDDR5 in MacBook. Apple was the first vendor to use 512-bit LPDDR5 and until today, no OEMs are able to do that.

iPhone Pro lineup going to get even more Pro features; one of them I think would be 12GB LPDDR6 support. With extra memory bandwidth, Apple going to add at least 2 more GPU cores thus pumping FP32 to around 3TF. Meanwhile, I suspect A18 would be like N3E version of A17 Pro featuring 6 GPU cores with 8GB LPDDR5.

Based on release timings, the rumored 8G4 (supposedly launch at Q4 2024, same timeline as iPhone 16) supporting LPDDR6 might be true cause 8G4's GPU performance is rumored to be around 35% faster. Guess what, A18 Pro's GPU also rumored to be around 38% faster.

Another advantage of LPDDR6 is cost and capacity. Most of the DUV-based LPDDR5/X die is using 16Gb (2GB) die size, thus in order to have 12GB, the memory chip have to use up to 6 dies. With EUV-based LPDDR6, 24Gb (3GB) is the standard thus Samsung only needs 4 dies to build a 12GB memory chip. It seems Apple going to use even bigger die size of 32Gb (4GB), that's means Samsung only needs 3 dies to build 12GB LPDDR6 memory chip. And that's how Apple able to fit 256GB memory in upcoming M4 Max. More on M4 series later.

 

[FlameTail]

As expected, two persons are using past history try to speculate Apple will be late to adopt LPDDR6. And they are ignoring that how aggressive Apple is adopting LPDDR5 in MacBook

LPDDR5 went mainstream in smartphones in early 2020, when many Android phones like Galaxy S20 series adopted it.

Meanwhile, Apple M1 which came in 2020Q4 stuck to LPDDR4X. Apple A14 Bionic (2020Q3) and A15 Bionic (2021Q3) also stuck to LPDDR4X.

It was only with M1 Pro and M1 Max in 2021Q4, that Apple finally adopted LPDDR5.

That is almost 2 years after LPDDR5 first debuted in smartphones.

Apple was the first vendor to use 512-bit LPDDR5 and until today, no OEMs are able to do that.

Bus width and LPDDR version have nothing to with each other. Other OEMs haven't done 512 bit coz it's not worth it for them.

I am willing to defend this hill till the end. Apple will not be early in adopting LPDDR6.

 

[Doug S]

As expected, two persons are using past history try to speculate Apple will be late to adopt LPDDR6.

And you have offered zero evidence for why Apple would change past history, aside from your bizarre obsession with LPDDR6 as apparently the greatest upgrade in the history of computing.

 

[Tigerick]

M4 Series with LPDDR6: Biggest Memory Upgrade

Gurman mentioned upcoming M4 series going to support up to 512GB. As shown in the table below, Apple going to use highest memory density (32Gb - 4GB) in order to support 256GB in upcoming M4 Max. The memory has to be LPDDR6 that build by EUV machine. M4 Max has doubled the memory size compared to M3 Max which support up to 128GB LPDDR5. It is considered biggest memory size upgrade cause the next upgrade will only happen in 2029 if my prediction is correct...

The M4 chips will be built on the same 3-nanometer process as the M3 chips, but Apple supplier TSMC will likely use an improved version of the 3nm process for boosts in performance and power efficiency.

@Doug S @FlameTail You two seems missing what Gurman wrote about M4 series going to be built by same 3-nanometer as M3 chips. Clearly it means N3B process which I have been saying for years. And you two also keep mentioned that M4 going to be build by N3E process. Or do you still think so? Do you want to wait until M4 launching then realize the mistake or admit now?

M-Max Series	Date	Max Memory	M-Ultra Max Memory
M1 Max	Q4 2021	64GB 512-bit LPDDR5	128GB 1024-bit LPDDR5
M2 Max	Q1 2023	96GB 512-bit LPDDR5	192GB 1024-bit LPDDR5
M3 Max	Q4 2023	128GB 512-bit LPDDR5	NA
M4 Max	Q4 2024 ?	256GB 384-bit LPDDR6	512GB 768-bit LPDDR6
	2029	384GB 384-bit LPDDR6	768GB 768-bit LPDDR6

 

[ToTTenTranz]

LPDDR5X actually goes all the way up to 9600MT/s and there's also LPDDR5T that starts at 9600MT/s.

It doesn't look like LPDDR6 is arriving anytime soon, considering JEDEC hasn't even approved of any spec yet.

 

[Doug S]

@Doug S @FlameTail You two seems missing what Gurman wrote about M4 series going to be built by same 3-nanometer as M3 chips. Clearly it means N3B process which I have been saying for years. And you two also keep mentioned that M4 going to be build by N3E process. Or do you still think so? Do you want to wait until M4 launching then realize the mistake or admit now?

I'll wait for M4 to launch and enjoy you eating crow. It will be N3E.

 

[SpudLobby]

I'll wait for M4 to launch and enjoy you eating crow. It will be N3E.

Yeah. I mean would it be anything other than N3E?

 

[Doug S]

Yeah. I mean would it be anything other than N3E?

He thinks it will be N3B, for what reason I'm not sure. I mean why wouldn't Apple want a better process that's faster and lower power? N3B is slightly denser, but Apple doesn't care about a few percent of density.