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

[Tigerick]

LPDDR6-12800 @ Q3-2024 ?

According to Samsung:

Next up is the LPDDR6, the next-generation JEDEC standard low-power DRAM, and it is no surprise that Samsung is leading the industry in defining the specification.

It seems JEDEC has approved the LPDDR6 standard (Update: Q3 2024) then Hynix and Samsung are going to mass produce LPDDR6. Micron is surprisingly not listing LPDDR6 in their roadmap...

Memory Types	Speed	64-Bit	128-Bit	192-Bit	256-Bit	384-Bit	512-Bit
LPDDR5	6400	51 GB/s	102 GB/s	154 GB/s	205 GB/s	360 GB/s	410 GB/s
LPDDR5X	8533	68 GB/s	136 GB/s	205 GB/s	273 GB/s	410 GB/s	546 GB/s
LPDDR6	12800	102 GB/s	205 GB/s	307 GB/s	410 GB/s	614 GB/s	819 GB/s
	24Gb	12 GB	24 GB	36 GB	48 GB
	14400	115 GB/s	230 GB/s	346 GB/s	461 GB/s	691 GB/s	922 GB/s
	32Gb	16 GB	32 GB	48 GB	64 GB
LPDDR6X	16000	128 GB/s	256 GB/s	384 GB/s	512 GB/s	768 GB/s	1024 GB/s

• 64-bit LPDDR6 most likely going to be used on future smartphone. Based on timing, Qualcomm's Snapdragon 8G4 which is manufactured by N3E going to become first OEMs to employ LPDDR6. Of course, Mediatek should follow closely...
• 64-bit and 128-bit LPDDR6 are going to become defacto choice of mainstream notebook computer.
• When Apple announced M3 Pro with 192-bit LPDDR5 support, I wonder what is Tim cooking? I think 192-bit LPDDR6 going to become next generation higher performance tier. Due to wider bandwidth gap, Apple will likely use 192-bit LPDDR6 for future M5 Pro while reserving 288-bit and 384-bit LPDDR6 for flagship model, namely M5 Max/Ultra.

DDR6-9600 @ FY2028 ?

Unfortunately, there may be some delays on DDR6 DIMM/SO-DIMM according to Samsung:

The roadmap calls for yearly DDR5 speed upgrades of up to 8800Mbps by the year 2027 with 9600Mbps DDR6 projected to be available by the year of 2028 with even higher speed rates.

Luckily, CPUs are not memory bound atm. That's mean even with doubling core counts,128-bit DDR5-8800 should have enough memory bandwidth, we shall see...

Memory Types	Speed	64-Bit	128-Bit
DDR4 (16-core Zen 3)	3200	26 GB/s	51 GB/s
DDR5 (16-core Zen 4)	5200	42 GB/s	83 GB/s
DDR5 (Arrow Lake-S)	6400	51 GB/s	102 GB/s
DDR5	7200	58 GB/s	115 GB/s
DDR5 (Nova Lake-S)	8000	64 GB/s	128 GB/s
DDR5	8800	70 GB/s	141 GB/s
DDR6	9600	77 GB/s	154 GB/s

LPDDR5/DDR5 Announcement's Timeline

1. At an online event on October 18, 2021, Apple announced their first MacBook Pro14/16 with M1 Pro/Max supporting latest LPDDR5-6400.
2. Few days later, Intel announced Alder Lake which can support both DDR4/DDR5 memory.
3. At CES 2022, AMD announced the Ryzen 6000 mobile series supporting DDR5/LPDDR5 only.

Apple, Intel and AMD are launching all their latest generation of CPU/APU within same timeframe. Based on timings, they should be working on the design three years ago, ie FY2019.That is the year JEDEC officially announced LPDDR5-6400 standard.

DDR5-6400 standard was announced in July 2020; one and half years later. That's mean Intel and AMD are working on DDR5 before official standard being announced.



Upcoming SoC with 64-bit LPDDR6: Redefine Value PC Segment

	MacBook Air 13		ARM 64-bit MC		Surface Go 3	Inspiron 14
Launch Price	$999				$730	$499
EOL Price	$699				$600	$280
Node	TSMC N5		Intel 3 / TSMC N3E		Intel 14nm	TSMC N7
CPU	M1				i3-10100Y	8cx Gen2
Core Counts	4P + 4E		2P + 6E / 4P + 4E		2P / 4T	4P + 4E
GB6 1T	2354				1211	1093
GB6 MT	8404				1522	3777
Memory	8GB LPDDR4x		12GB LPDDR6		8GB LPDDR3	8GB LPDDR4x
Memory BW	68 GB/s		102 GB/s		33.3 GB/s	68 GB/s
GPU	8 cores				UHD 615 24EU	Adreno 690
FP32	2.6 TF		~ 4 TF		0.4 TF	~ 2 TF
AIE	11 TOPS		~ 40 TOPS ?		NA	9 TOPS
Features	Unibody Retina 2560x1600 15-Hour Wireless Web				LTE 3:2 Multi-Touch	5G Option

 

[Tigerick]

AMD: ~2025 for Sarlak6

Memory Types	Speed	64-Bit	128-Bit	192-Bit	256-Bit
LPDDR5(X)	5500-8533	Mendocino	Hawk Point	Sarlak	Sarlak
		Sonoma Valley	Kraken Point
			Strix Point
LPDDR6	12800	ARM ?	ARM ?	Sarlak6	Sarlak6

Year	Sarlak's CPU	Sarlak's GPU	Memory	Memory BW	Estimated CPU's BW	GPU BW		RX 7600M (RDNA3, N6)
2025	16 x Zen 5	RDNA3+ 40CU	256-bit LPDDR5X-8533	273 GB/s	42 GB/s	231 GB/s		256 GB/s
	2 X N4P	N3E	24 - 32 GB					28 CU @ 2.41GHz
								RX 8700M (RDNA4, N4P)
2026	16 x Zen 6	RDNA5 40CU ?	256-bit LPDDR6-12800	410 GB/s	42 GB/s	368 GB/s		400 GB/s ?
	N3P	N3E	36 - 48 GB	+ 50% + 200%				40 CU @ ?

CPU	Model	Cores	Node	tCPU	L3 Cache	TDP		GB6 Single	Points Per Watt		CB 2024 Single	Points Per Watt		GB6 Multi	Points Per Watt		CB 2024 Multi	Points Per Watt
Zen 2	Ryzen 9 3950X	16	N7	2	64 MB	105 W		1705	16.2		74	0.70		10691	101.8		1325	12.6
Zen 3	Ryzen 9 5950X	16	N7	2	64 MB	105 W		2185	20.8		95	0.90		12246	116.6		1494	14.2
Zen 4	Ryzen 9 7950X	16	N5	2	64 MB	170 W		2936	17.3		125	0.74		19293	113.5		2142	12.6
Zen 5	Ryzen 9 9950X	16	N4P	2	64 MB	?		3500	20.6		170	0.82		24000	141.2		3000	14.1
Zen 6	?	16	N3P	1	32 MB ?	?
Zen 6C	?	32	N2	1	64 MB ?	?

 

[Tigerick]

Qualcomm: ~Q4 2025 for X Elite G2

	X Elite G1	X Elite G2		Hawk Point	ARM's 64-bit	AMD's ARM	Strix Point
Node	N4P	N3P		N4	?	N3E / N3P	N4P
SoC	Monolithics	Monolithics		Monolithics	Monolithics	Monolithics	Monolithics
Memory Types	LPDDR5X-8533	LPDDR6-12800		LPDDR5X-7500	LPDDR6-12800	LPDDR6-12800	LPDDR5X-8533
Memory Bus	128-bit	128-bit		128-bit	64-bit	128-bit	128-bit
Memory BW	136 GB/s	205 GB/s		120 GB/s	102 GB/s	205 GB/s	136 GB/s
CPU	12P Oryon	Pegasus ?		8P Zen4	6-8 P ?	12-16 P ?	4 x Zen5 + 8 x Zen5c
GPU	Adreno	Adreno		RDNA3	RDNA3+	RDNA3+	RDNA3+
RT	NA	Y ?		?	Y	Y	Y
SM/CU				12 CU	12 CU ?	24 CU ?	16 CU
FP32	4.6 TF	?		4.3 TF	~ 4.6 TF	~ 9.2 TF	~ 6.1 TF
AIE (TOPS)	45	?		16	?	?	50

 

[Tigerick]

Apple: Q4-2024 for M4 series ?

Memory Density 16Gb (2GB)	M3	M3 Pro	M3 Max	M3 Max
LPDDR5 Memory Bus	128-bit	192-bit	384-bit	512-bit
Memory BW	100 GB/s	150 GB/s	308 GB/s	410 GB/s
Memory Chips	2 pcs	3 pcs	3 pcs	4 pcs
x2 = 4GB (S: 32-Gb)	8 GB
x3 = 6GB (M: 48-Gb)		18 GB
x4 = 8GB (S: 64-Gb)	16 GB
x6 = 12GB (S: 96-Gb)	24 GB	36 GB	36 GB	48 GB
x8 = 16GB (S: 128-Gb)				64 GB
x16 = 32GB (256-Gb)			96 GB	128 GB

Memory Density 32Gb (4GB)	M4	M4 Pro	M4 Max	M4 Max
LPDDR6 Memory Bus	96-bit	192-bit	288-bit	384-bit
Memory BW	150 GB/S	307 GB/s	461 GB/s	614 GB/s
BW+	+ 50 GB/s	+ 150 GB/s	+ 153 GB/s	+ 214 GB/s
Memory Chips	1 pc	2 pcs	3 pcs	4 pcs
x3 = 12GB	12 GB	24 GB	36 GB	48 GB
x5 = 20GB	20 GB	40 GB		80 GB
x7 = 28GB	28 GB	56 GB
x8 = 32GB			96 GB	128 GB
x12 = 48GB				192 GB
x16 = 64GB				256 GB

Apple Silicon	Date	Node	Transistors	SLC	Memory	CPU	GPU
M3	Q4 2023	N3B	25 B	8 MB	8GB 128-bit LPDDR5	4P + 4E	10 RT
M4	2025 ?	N3B		12 MB ?	12GB 96-bit LPDDR6	?	?
M3 Pro	Q4 2023	N3B	37 B	12 MB	18GB 192-bit LPDDR5	6P + 6E	18 RT
M4 Pro	2025 ?	N3B		24 MB ?	24GB 192-bit LPDDR6	?	?
M3 Max	Q4 2023	N3B	92 B	48 MB	36GB 384-bit LPDDR5	10P + 4E	30 RT
M4 Max	2025 ?	N3B		36 MB ?	36GB 288-bit LPDDR6	?	?
M3 Max	Q4 2023	N3B	92 B	64 MB	48GB 512-bit LPDDR5	12P + 4E	40 RT
M4 Max	2025 ?	N3B		48 MB ?	48GB 384-bit LPDDR6	?	?

	Clock Speed	GB6 1T	Points Per Clock	GB6 MT	GPU cores	Metal	Points Per Core
M1	3.2 GHz	2354	736	8404	8	32104	4013
M2	3.49 GHz	2634	754	9753	10	45656	4566
M3	4.06 GHz	3084	760	11564	10	47372	4737

	Memory	Memory BW	GPU cores	Metal	Points Per Core
M1 Pro	256-bit	200 GB/s	16	66258	4141
M2 Pro	256-bit	200 GB/s	19	81551	4292
M3 Pro	192-bit	150 GB/s	18	78218	4345
M4 Pro	192-bit LPDDR6	300 GB/s	?	?	?

	CPU cores	Clock Speed	GB6 1T	Points Per Clock	GB6 MT	GPU cores	Metal	Points Per Core
M1 Max	8P + 2E	3.2 GHz	2381	744	12233	32	113388	3543
M2 Max	8P + 4E	3.66 GHz	2738	748	14518	38	137677	3623
M3 Max 384-Bit	10P + 4E	4.06 GHz	3123	769	19001	30	123618	4121
M3 Max	12P + 4E	4.06 GHz	3125	770	20984	40	154245	3856

 

[Tigerick]

Intel: ~Q4 2026 for Nova Lake-H

Desktop CPU

1. Arrow Lake-S 8P+16E (N3B): Q4 2024
2. Arrow Lake-S 8P+32E (N3B): Q4 2025
3. Nova Lake-S 16P+32E (18A-P/N2): Q4 2026 comes with 128-bit DDR5-8000

Mobile APU:

• Nova Lake-H 8P+16E (Intel 18A-P ?) + Celestial (N3P ?): Q4 2026

	Lunar Lake	? Lake	Meteor Lake ICU7-164U	Alder Lake-N i3-N300		Panther Lake-H	Nova Lake-H	Nova Lake-S / HX
Date	Q4 2024	?	Q4 2023	Q1 2023		Q4 2025	Q4 2026 ?	Q4 2026
Node	N3B + N6	Intel 3 ?	Intel 4 + N5 + N6 + N6	Intel 7		18A + N3E	18A-P + N3E ?	N2 + Intel 3
TDP	17 - 30 W	?	9 - 30 W	7 W
Memory	128-bit LPDDR5X-8533	64-bit LPDDR6-12800	128-bit LPDDR5X-6400	64-bit LPDDR5-4800		128-bit LPDDR5X-8533	128-bit LPDDR6-12800	128-bit DDR5-8000
Memory BW	136 GB/s	102 GB/s	102 GB/s	38.4 GB/s		136 GB/s	205 GB/s	128 GB/s
CPU	4P + 4E	?	2P + 8E	8E / 8T		4P + 8E	8P + 16E ?	16P + 32E ?
GPU	ARC 8 XE	?	4 XE	UHD 32EU		ARC 12 XE	ARC 12 XE ?	4 XE
Clock Speed	1.85 GHz	?	1.8 GHz	1.25 GHz		2.5 GHz	3 GHz	?
FP32	3.8 TF	?	1.8 TF	0.6 TF		7.7 TF	9.2 TF	?

 

[igor_kavinski]

DDR5 not going beyond 8800 MT/s is disappointing.

DDR6, whenever it gets released, is going to have terrible CAS latency compared to DDR5 and the same old crap will ensue where you have to wait a year or two for the newer RAM standard to get fast enough to leave the older standard behind. This, in my opinion, is the great RAM speed scam.

Why do they even release a new generation of memory when the previous more mature standard already beats it at its fastest available speeds? DDR5-4800 vs. DDR4-4000 was a sick joke.

Why don't these "brilliant" engineers actually try to bring the latencies down to an acceptable level before proudly touting their new technologies? It's actually not their fault. They would never release something half baked into the world if it were up to them. It's the profit craving executives/bean counters calling the shots at these big companies. The bane of every techie's existence.

/rant

 

[itsmydamnation]

DDR5 not going beyond 8800 MT/s is disappointing.

DDR6, whenever it gets released, is going to have terrible CAS latency compared to DDR5 and the same old crap will ensue where you have to wait a year or two for the newer RAM standard to get fast enough to leave the older standard behind. This, in my opinion, is the great RAM speed scam.

Why do they even release a new generation of memory when the previous more mature standard already beats it at its fastest available speeds? DDR5-4800 vs. DDR4-4000 was a sick joke.

Why don't these "brilliant" engineers actually try to bring the latencies down to an acceptable level before proudly touting their new technologies? It's actually not their fault. They would never release something half baked into the world if it were up to them. It's the profit craving executives/bean counters calling the shots at these big companies. The bane of every techie's existence.

/rant

Because DRAM cells haven't improved in like 30 years. What your asking for will result in 0 investment in creating memory that delivers higher performance because the upfront investment before any type of return just exploded. Tech is diminishing return curves and the next curve always starts bellow the end of the last one.

Get over it.

 

[igor_kavinski]

What your asking for will result in 0 investment in creating memory that delivers higher performance

Problem is, they are creating higher bandwidth memory to feed more cores which is mostly a server sector problem.

Consumers don't really need that much bandwidth coz most of them don't have that many cores to feed.

Consumers need the lowest possible latency.

It's insulting to have to eat food cooked for "servers" when consumer CPUs far outnumber server CPUs.

One should never back down from speaking against injustice.

 

[SarahKerrigan]

Problem is, they are creating higher bandwidth memory to feed more cores which is mostly a server sector problem.

Consumers don't really need that much bandwidth coz most of them don't have that many cores to feed.

Consumers need the lowest possible latency.

It's insulting to have to eat food cooked for "servers" when consumer CPUs far outnumber server CPUs.

One should never back down from speaking against injustice.

Injustice! How dramatic.

 

[igor_kavinski]

Injustice! How dramatic.

Easy for you to make fun of the common man from your high horse.

The majority of tech consumers do not have access to fancy server hardware.

 

[poke01]

sigh but LPDDR6/X needs to be an option on desktop. PC handhelds and Win laptops/tabs are going to be even more great for gaming.

 

[itsmydamnation]

Problem is, they are creating higher bandwidth memory to feed more cores which is mostly a server sector problem.

Consumers don't really need that much bandwidth coz most of them don't have that many cores to feed.

Consumers need the lowest possible latency.

It's insulting to have to eat food cooked for "servers" when consumer CPUs far outnumber server CPUs.

One should never back down from speaking against injustice.

put your money where your mouth is then and invest in a dram replacement company then.

edit: your also ignoring the improvements that come from more concurrency at the same latency

 

[VirtualLarry]

The majority of tech consumers do not have access to fancy server hardware.

Zen Desktop is derived from Zen Server. It's not so bad.

 

[igor_kavinski]

Zen Desktop is derived from Zen Server. It's not so bad.

Talking memory here. Not CPU. No offense.

 

[SarahKerrigan]

Easy for you to make fun of the common man from your high horse.

The majority of tech consumers do not have access to fancy server hardware.

Hardly making fun of the common man or anyone else. If anything, I'm glad someone appreciates the hard work of the architects of several generations of RLDRAM and QDR SRAM.

The pricing, of course, is the less appealing part.

 

[FlameTail]

Qualcomm: ~Q2 2026 for X Elite G2

	X Elite G1	X Elite G2		nVidia	Panther Lake		AMD	Strix Point
Node	N4P	N3P		N3P	18A + 3+		N3P	N4P
SoC	Monolithics	Monolithics		Monolithics	2 Tiles		Monolithics	Monolithics
Memory Types	LPDDR5X-8533	LPDDR6-12800		LPDDR6-12800	LPDDR5X-8533		LPDDR6-12800	LPDDR5X-8533
Memory Bus	128-bit	128-bit		128-bit	128-bit		128-bit	128-bit
Memory BW	136 GB/s	205 GB/s		205 GB/s	136 GB/s		205 GB/s	136 GB/s
CPU	12P Oryon	Pegasus ?		?	4P+8E		?	4 x Zen5 + 8 x Zen5c
GPU	Adreno	Adreno		Blackwell ?	Arc		RDNA3+	RDNA3+
RT	NA	Y ?		Y	Y		Y	Y
FP32	4.6 TF	?		?	7.7 TF		?	~ 6.1 TF
AIE (TOPS)	45	?		?	50 ?		?	50

Hey where did you get the X Elite G2 launch date from?

 

[Tigerick]

Hey where did you get the X Elite G2 launch date from?

Estimated

 

[FlameTail]

Estimated

X Elite G2 coming in Q2 2026?

That is crazy. 2.5 years from X Elite announcement and 2 years from X Elite release?

That makes no sense. Bad estimate imo.

Intel and AMD will kill Qualcomm's renewed WoA effort, if Qualcomm releases the X Elite successor so long after.

 

[FlameTail]

Here's what I think:

2023Q4 : X Elite announced
2024Q2 : X Elite released
2025Q4: X2 Elite announced/released (estimate)
2026Q4 : X3 Elite announced/released (estimate)

Q4 is when Qualcomm holds their annual Snapdragon Summits. If X2 Elite is announced/released at 2025Q4 SS, that is 2 years after X Elite announcement, and 1.5 years after X Elite release. That sounds much more reasonable.

 

[Glo.]

Problem is, they are creating higher bandwidth memory to feed more cores which is mostly a server sector problem.

Consumers don't really need that much bandwidth coz most of them don't have that many cores to feed.

Consumers need the lowest possible latency.

It's insulting to have to eat food cooked for "servers" when consumer CPUs far outnumber server CPUs.

One should never back down from speaking against injustice.

For latency you will have to have large system level caches.

Thats the reality of future, even consumer products.

 

[FlameTail]

@Tigerick

The chart says 2026-LPDDR6, but you know what?

Assuming the Snapdragon X2 Elite launches in 2026-Q2 (or any other PC SoC for that matter), it's doubtful it will have LPDDR6, let alone the top spec LPDDR6-12800.

The reason is that initially LPDDR6 will be very expensive, limited in volume and won't reach the max speed of 12800.

2020 was the maiden year for LPDDR5, being shipped widely in flagship Android smartphones. But these LPDDR5-5500, and it was not until 2021 that the full spec LPDDR5-6400 shipped widely.

LPDDR6 will follow a similar path. If 2026 is the maiden year for LPDDR6, in that year it will be limited to mostly flagship Android phones (which are low volume), and will probably be LPDDR6-11000 or something.

Large volume of LPDDR6 and full speed 12800 will come in 2027.

 

[Tigerick]

DDR6 RAM: The next-gen memory standard, explained

Unsurprisingly, DDR6 memory will gradually replace the current DDR5 memory over the next few years, but what do we actually know about the upcoming memory standard? We'll tell you.

www.pcworld.com

Even though the article is dated two days ago, the content is actually translated from Germany website pcwelt.de about Tech Day 2021. Some errors are to be expected: -

• DDR6-12800 is actually meant LPDDR6-12800 as stated in my front page. Samsung actually said DDR6 DIMM is delayed beyond 2028. Therefore, what the article claimed: "The new DDR6 memory will celebrate its debut in the professional enterprise segment and is expected to be used for the first time with the server CPUs of the 2025 vintage" could be wrong. That might explain the expansion of 12-channel DDR5 to 16-channel in upcoming AMD EPYC server platform.
• My estimated LPDDR6X is rated at 16000 MT/s derived from Samsung roadmap. According to Samsung, they need advanced EUV (High-NA?) to hit the speed of 8Gbps, therefore don't expect LPDDR6X to hit the market soon.
• Calculations of memory bandwidth are wrong on DDR6/LPDDR6, you guys can check my table to find out each corresponding memory bandwidth with different memory bus.

Some worthy information below:

• Samsung has already announced that it is currently working on finalizing the DDR6 memory standard together with the other JEDEC members from the circle of DRAM and SoC manufacturers. According to the manufacturer, this is expected in the course of 2024, but no later than 2025.
• The chief engineer of the South Korean semiconductor manufacturer SK Hynix, on the other hand, expects a somewhat longer development phase and does not anticipate a broad market launch before the end of 2025.
• As with the switch from DDR4 to DDR5, the feature set of the new memory generation will once again be significantly expanded. All relevant DRAM chip manufacturers, such as Samsung, Micron, Nanya, and SK Hynix, have already revealed this.

While Samsung might be preparing to launch LPDDR6 in 2025, Hynix expected broad market launch of LPDDR6 by end of 2025 is aligned with what Samsung roadmap revealed. Of course, delay could happen but Q4 2025 is the projected launch windows of SoC with LPDDR6-12800...

 

[Tigerick]

Large volume of LPDDR6 and full speed 12800 will come in 2027.

Do you still think about 2027? How about I show you the leaks below:



Last time you shown me the slide, I thought SD 8 Gen 5 should use N3P, but what if the Korean guy is correct? Then the slide below might explain the LPDDR6 version of SD coming in 2025 which means Q4....

 

[Tigerick]

As I learned more about Samsung roadmap, I think I know why Micron might not be able to produce LPDDR6 memory. LPDDR6 needs to hit 6.4Gbps speed in order to arrive at 12800. They need EUV machine to make it. Currently only Samsung and Hynix have enough EUV machines to mass produce it.

If you are familar with Micron, they are producing current memory with DUV (older generation of machinery), thus they are only able to hit 4.8Gbps with LPDDR5X. Unless Micron is migrating to EUV, they won't be able to make LPDDR6. Yeah, that's mean Micron has dropped out of race.

As for people wondering about the BOM, they should at least learn about EUV advantage. Yeah, the machine is expensive so does the wafer, but the advantages are smaller die size with bigger capacity and higher clock speed. Especially bundle with standard 64-bit memory bus, the PCB would be cheaper too...

 

[FlameTail]

Do you still think about 2027? How about I show you the leaks below:

View attachment 94587

Last time you shown me the slide, I thought SD 8 Gen 5 should use N3P, but what if the Korean guy is correct? Then the slide below might explain the LPDDR6 version of SD coming in 2025 which means Q4....

View attachment 94588

Dubious leaks.