Track Hyper | DDR6 New Standard Strikes: Boosting AI PC Performance

Author: Zhou Yuan / Wall Street News

DDR6 hardware standard is starting to emerge. This is only 4 years after the standardization of DDR5 memory.

According to information from JEDEC (Solid State Technology Association), DDR6 (including LPDDR6) will explicitly replace the long-used SO-DIMM and DIMM memory standards with CAMM2.

This is particularly important now.

Why? Because the new category of AI PCs will have much higher requirements for memory speed, power consumption, and bandwidth compared to traditional PCs. The minimum speed of LPDDR6 is 10.667Gbps, equivalent to the highest speed of LPDDR5x, far exceeding LPDDR5 at 6.7Gbps. The maximum speed of LPDDR6 can reach 14.4Gbps.

Due to the new 24-bit wide channel design of LPDDR6, the memory bandwidth can reach up to 38.4GB/s, much higher than the current LPDDR5 standard. With higher data transfer rates and bandwidth, LPDDR6 can better meet the future AI computing demands, performing far better on AI PCs than existing memory standards.

More importantly, CAMM2 not only supports higher data transfer rates and lower memory latency, but also meets the requirements of AI computing, thus outperforming the current LPDDR5 SO-DIMM/DIMM standards in terms of performance; at the same time, CAMM2 also supports modular design.

This design will change the non-upgradable nature of LPDDR5. What is LPDDR5? It is a low-power memory mainly designed for platforms such as smartphones and mobile PCs. This standard memory makes more compromises in terms of power consumption and hardware volume, with higher integration.

Although PC memory can be disassembled, this is only for conventional memory, mainly for desktop PCs. A considerable number of ultra-thin laptops, for the sake of "thinness" and portability, use LPDDR5 memory that is soldered onto the motherboard. This makes it impossible to "add memory" to improve overall performance.

Therefore, current ultra-thin laptops face the issue of being unable to upgrade in the future. Not to mention meeting the performance requirements of future AI PCs, even improving current performance is difficult due to the inability to upgrade, making such expectations hard to achieve. In such cases, either buy a new machine, which may still face the issue of being unable to upgrade in the future, or reluctantly continue using it, which is quite unsatisfactory.

Although the CAMM2 new standard continues the high integration design of the previous SO-DIMM/DIMM, this new standard has adopted modular design from the beginning of its formulation, so CAMM2 can use universal interfaces and be self-replaceable.

With the support of new generation integration technology and modular design standards, the thickness of LPDDR6 adopting the CAMM2 standard is even lower than the mainstream SO-DIMM standard LPDDR5x/LPDDR5 currently used In terms of physical form, LPDDR6 is no longer in the previous "strip" shape, but has become a "slice" shape like a PCB board. Due to the ultra-thin characteristics of memory chips, it is similar to soldering ultra-thin memory chips and controllers on a PCB board, making the overall physical form appear as a thin board.

CAMM2 is a new memory standard that also supports the DDR6 standard, suitable for large PC devices such as desktop PCs. This is the first time since the introduction of the DIMM memory standard in 1993 that the PC market has truly welcomed a brand new standard.

However, based on the true product iteration history of the industry from the DDR5 memory standard, it is likely that it will take no less than 5 years for DDR6 to become mainstream.

Because DDR5 has not yet surpassed DDR4 (in terms of market share) since establishing the standard until now. The industry predicts that by 2025-2026, DDR5 may surpass DDR4 in market share. This is still under the condition that DDR5 and DDR4 memory standards are consistent, both using SO-DIMM or DIMM standards.

Currently, the upstream memory manufacturers like Samsung, SK Hynix, Micron, and China's Loongson Technology have all supported the CAMM2 standard; downstream OEMs, in 2023, Dell used CAMM2 memory boards in its enterprise product line. There are reports that Lenovo will follow suit in 2024 and gradually expand it to the consumer line.

On April 25th this year, Lenovo Group launched the ThinkPad P1 Gen 7 personal high-performance mobile workstation, belonging to the new category of AI PCs.

This model is equipped with Intel Core Ultra processors and NVIDIA RTX Ada workstation graphics cards; it also uses the industry's first standard low-power compressed additional memory module from Micron (Crucial LPDDR5x-7467MT/s LPCAMM2 memory module): with transitional LPDDR5x-7467MT/s.

Despite having transitional properties, the performance and physical elements (such as size and weight) of this memory are far superior to traditional LPDDR5x using the mainstream SO-DIMM/DIMM standards.

From a technical perspective, this memory solution not only provides the strongest energy-efficient modular memory performance currently available but also can reach up to 64GB, with a 1.3 times speed increase.

Compared to DDR5 using SO-DIMM/DIMM, CAMM2 LPDDR5x reduces power consumption by 58%, saves 64% of space, provides higher bandwidth and dual-channel support through a single module, making it an ideal high-performance memory solution for mobile workstations and AI PC workloads Due to its highly space-saving design, the overall dimensions of the ThinkPad P1 Gen 7 are 354.40 x 241.20 x 17.05 mm, yet the weight is kept at around 1.82kg, making it remarkably slim and lightweight for a 16-inch laptop with a discrete graphics card.

However, this might lack a tangible sense of comparison.

Let's directly compare it with the newly released ThinkPad P16v i Gen 2 by Lenovo. Both models have the same chipset and graphics card, but the P16v i Gen 2 uses the traditional SO-DIMM standard LPDDR5 memory, resulting in overall dimensions of 365 x 262 x 24.66mm and a starting weight of 2.22kg.

See the difference? Just by using memory of different generational standards, these two super personal workstations have a size difference of 10.6mm x 20.8mm x 7.61mm, a thickness difference of 7.61mm, and a weight difference of 0.4kg.

Of course, this is a rather rough comparison. With the stacked dual-slot design, DIMM memory modules will increase the thickness by at least 5mm+ compared to CAMM2.

Apart from the differences in size and weight, performance is the biggest advantage of the CAMM2 standard. However, the current cost difference is also hard to ignore, considering that models with the same hardware configuration but different memory standards have a retail price difference of at least $340 (approximately RMB 2500).

Earlier, it was mentioned that this is a transitional memory product. In reality, true CAMM2 standard LPDDR6 memory, for example, in a 32GB specification, costs around $500, which is 5 times the price of LPDDR5 (SO-DIMM/DIMM) memory.

Nevertheless, if the AI PC market rapidly expands, the replacement of SO-DIMM/DIMM memory standards by CAMM2 is likely to happen at a much faster pace than before