Soldered vs. on-package: not the same thing
Two different designs get lumped together under "soldered RAM," and the distinction matters.
- Soldered memory is mounted directly on the motherboard. Buyers generally can't upgrade it.
- Memory on package puts the memory chips on the same substrate as the processor. Intel did this with Lunar Lake (Core Ultra Series 2).
- Unified memory, Apple's term, is a pool shared by the CPU and GPU.
In every case you pick a capacity at checkout and live with it. The physical layouts and cost structures differ, though. Memory speed, bus width and controller design also shape performance, so shorter traces are only part of the story. Comparing a soldered LPDDR configuration with a socketed SO-DIMM one usually compares several design choices at once. It doesn't isolate the effect of solder.
The performance case: bandwidth
The strongest argument is bandwidth, which matters most for integrated graphics. An iGPU has no dedicated video memory, so textures, frames and shader data all compete for the same system RAM bandwidth.
AMD's Ryzen AI Max+ 395 is the clearest example. Several spec listings agree on its configuration:
- It pairs 16 Zen 5 cores with a 40-CU Radeon 8060S GPU and up to 128GB of LPDDR5X-8000 on a 256-bit bus, for 256 GB/s.
- The 256-bit interface is twice as wide as on normal notebook processors.
- The memory is soldered and cannot be upgraded.
That width is the real advantage. Socketed SO-DIMM laptops typically use a narrower dual-channel arrangement. The MakeUseOf piece puts typical SO-DIMM bandwidth at roughly 90GB/s, but I couldn't verify that figure independently.
Two caveats apply:
- 256GB/s is a theoretical peak. One local-AI spec aggregator lists about 215GB/s measured against that theoretical number. Treat it as a ceiling, not a guarantee.
- The "RTX 4070-class" graphics claim is unverified. I found no benchmark that supports it. The answer depends on power limits, the specific laptop and the workload. One hardware roundup describes the chip only as "roughly in the ballpark" of a mobile RTX part. That is a much softer claim.
Apple's numbers are easier to pin down. Apple's M5 MacBook Air announcement cites faster unified memory with 153GB/s of bandwidth, a 28 percent improvement over M4. The article's 9,600 MT/s figure for the M5 doesn't appear in that official material, so I'd leave it out.
Intel's Lunar Lake: the instructive case
Lunar Lake took integration further than almost anyone. Intel put LPDDR5X on the processor package. Intel's materials, per the research brief, claim about 40% lower power for moving data and a smaller board footprint. Those are Intel's own claims, not independent measurements.
The cost side is documented too. PC Gamer's January 2026 report quotes Intel CFO David Zinsner on the economics:
- Intel said it had enough memory stockpiled to keep selling Lunar Lake, having been "relatively aggressive" in securing it early.
- Zinsner also said Lunar Lake's margins are low because the memory is packaged. Extra chips, extra packaging steps and securing memory all add cost compared with other chips.
- Pat Gelsinger, Intel's former CEO, had already called on-package memory "a one-off with Lunar Lake." He said Panther Lake, Nova Lake and their successors would keep volume memory off-package.
So the MakeUseOf piece is directionally right that cost pushed Intel back. But "specifically negatively affected Intel's bottom line" overreaches. The evidence shows reduced gross margins on that product, not a verified company-wide financial outcome. It also doesn't show the idea failed technically. It shows that on-package memory is a strategic trade-off, not a free lunch.
The memory crunch cuts both ways. A chip with memory built in exposes its maker to memory prices directly. That's the exposure Zinsner described.
Space, thinness and cooling
Sockets take room. Removing them frees board area for a bigger battery, better cooling or a thinner chassis. The MakeUseOf article attributes a 250 square millimeter board-area saving to Intel's Hot Chips presentation. I couldn't confirm that figure in the sources I could access, so attribute it to Intel if you repeat it. The general principle is sound and widely understood in the industry.
Cost savings are less certain. The article quotes an HP executive saying a socket costs more than soldering the RAM, via Digital Trends. I couldn't check the original interview. It's a plausible claim, but it's one executive's statement with no numbers, not an established cost comparison.
The upgradeability counterargument
Upgradeable machines haven't gone away. Framework's current Laptop 12 specs show that:
- Its Intel Core Series 3 version has a single SO-DIMM slot supporting up to 64GB of DDR5-6400.
- The same page lists pre-built options of up to 48GB DDR5-5600, or you can bring your own memory.
- Framework markets the machine on repairability, and the spec page describes it as easy to repair with parts sold in its marketplace.
That's a real alternative for people who want to start small and add RAM later. It also shows that fixed memory isn't the only design left. The trade-off is that a socketed design gives up the wide, fast bus that the integrated designs above rely on.
Practical buying guidance for Windows users
- Check the exact model's memory spec. Look for "soldered," "LPDDR5X" or "on-package" in the spec sheet, and check whether any SO-DIMM slot exists. Many thin-and-light laptops list soldered memory plus a maximum capacity.
- Buy for the whole ownership period. With fixed RAM, the capacity you choose at purchase is the capacity you'll have for the life of the laptop. Think about where your workloads will be in three to five years.
- Weigh your workload. Virtual machines, video editing, 3D work and local AI models use memory heavily. For those, either buy a high-capacity configuration up front or choose a machine you can upgrade.
- For integrated graphics, look at bandwidth, not just capacity. Memory speed and bus width can matter as much as gigabytes. Check the memory type and bus width where the manufacturer publishes them.
- Treat "enough for the foreseeable future" as personal. The article says 32GB is enough for its author and 16GB for most people. That's a workload judgment, not a verified rule. The article's 2015-versus-today table of typical RAM sizes also cites no data, so I'd treat it as the author's impression.
Verdict
"Soldered RAM is the best laptop innovation" is a bold headline, and a more defensible version is narrower. Memory integration, whether soldered or on package, enables high bandwidth, efficient data movement, compact boards and strong integrated graphics. In return you accept a fixed capacity, less repairability and more exposure to memory pricing.
For a thin travel laptop that you'll replace in a few years, the trade is probably fine. For a machine you want to keep for a decade, or one that might hit memory-hungry work, an upgradeable laptop still makes sense. The MakeUseOf piece also takes its own shot at AI-driven memory demand, but that aside rests on no cited evidence, so I haven't repeated it.