My laptop stopped keeping up: how I am choosing the next one

I took eleven products to release on an ASUS Zenbook 14 OLED (UM3402YA). It is thin, quiet, and has a genuinely good OLED panel, and I still think it was a sound purchase. But it stopped keeping up, and this time the problem is not that the hardware aged. It hit a ceiling the manufacturer built into it.

Below is the breakdown: what my work actually runs out of, what changed in the 2026 platforms, and the criteria I am using to pick a replacement. Plus a section on the numbers I checked that did not hold up.

What a 2022 ultrabook runs out of

The UM3402YA has two limits, and neither can be removed.

The memory is soldered, and the ceiling is 16GB of LPDDR4X-4266. That is not a configuration you can top up later, it is the platform maximum. Dual-channel LPDDR4X-4266 gives roughly 68GB/s of bandwidth. For comparison, the base Apple M5 delivers 153GB/s, more than twice as much.

The Ryzen 5 5625U is Zen 3 in an ultrabook power envelope. Six cores, twelve threads, 2.3GHz base and up to 4.3GHz boost, 16MB of L3, 7nm, rated at 15 watts. It handles an editor and a small build without complaint. It does not handle several heavy processes at once: in a 15-watt envelope the boost does not hold, and sustained load pulls clocks back towards base.

Here is what that looks like in practice. A normal day for me means a dev server, a local database, a browser with twenty-odd documentation tabs, an AI assistant in the editor, and a Python data run every so often. None of those is a problem on its own. Together they stop fitting into 16GB, and the system starts swapping.

Swapping is where everything falls apart. While data sits in RAM, the processor waits nanoseconds for it. Once the system pushes part of it onto the SSD, the wait grows by orders of magnitude, and this drive is PCIe 3.0, not the quickest. From the outside it does not read as “slow”, it reads as “frozen”: the interface stops responding for seconds at a time. The processor is barely loaded during those seconds. It is waiting.

That produced the criterion I now select on: for my work, memory capacity and speed matter more than peak CPU clock. Most reviews rank those the other way around.

What actually changed by 2026

Apple M5: the core count is not the story

Compare M4 and M5 by core count and there is almost nothing to see: both base chips go up to a 10-core CPU. The arrangement is the interesting part.

M5 is built on third-generation 3-nanometre process. The CPU is six efficiency cores and up to four performance cores. Apple states up to 15 percent faster multithreaded performance over M4, and that is the only official CPU performance figure for the generation.

The real change is in the graphics. In M4, machine learning ran on a separate block, the Neural Engine. In M5, Apple put a Neural Accelerator into each of the ten GPU cores and claims over 4x the peak GPU compute performance for AI compared to M4. Separately: up to 30 percent faster graphics, and a 45 percent uplift in apps using ray tracing. The Neural Engine is still there, at 16 cores.

For my work the third number matters most: unified memory bandwidth went from 120GB/s to 153GB/s, close to 30 percent. Memory on the base M5 tops out at 32GB.

M5 Pro and M5 Max: this is where super cores live

Precision matters here, because popular write-ups get it wrong. Apple applies the term “super cores” not to the base M5 but to the higher chips. M5 Pro and M5 Max carry an 18-core CPU: 6 super cores and 12 performance cores.

The rest of the specification diverges too:

• M5 Pro: up to a 20-core GPU, up to 64GB of memory, 307GB/s, up to 30 percent faster multithreaded than M4 Pro.
• M5 Max: up to a 40-core GPU, up to 128GB of memory, 614GB/s, up to 15 percent faster multithreaded than M4 Max.

The gap between Pro and Max is not a rounding difference: double the memory and double the bandwidth. If you are reading a table that lists 128GB against M5 Pro, the table is wrong.

x86: AMD and Intel

AMD Ryzen AI 9 365 (Strix Point) is 10 cores, Radeon 880M graphics on RDNA 3.5, and an XDNA 2 neural engine rated at 50 TOPS. It supports DDR5-5600 and LPDDR5X-8000, with peak memory bandwidth of 89.6GB/s. That figure, not the “around 120” that gets repeated.

Intel Core Ultra Series 3 (Panther Lake) ships on the 18A process, up to 16 cores, with a fifth-generation NPU. There is no single bandwidth figure, and that matters when buying: entry SKUs cap at LPDDR5X-6800, mid-range at 8533, and the top parts support 9600 MT/s and clear 150GB/s. The spread inside one product line is wider than the gap between Apple generations.

Qualcomm Snapdragon X2 Elite on third-generation Oryon: 152GB/s on the regular parts and 228GB/s on the Extreme, where memory runs across twelve channels and reaches 128GB on package.

Platform comparison

Reading that table against your own work

If the load is data and Python. Look at capacity and bandwidth together, in that order. A dataset that fits entirely in RAM is processed quickly even on a narrow bus. A dataset that does not fit destroys performance regardless of the processor. Which is why 64GB on a Ryzen with an 89.6GB/s bus will be more practical for many jobs than 32GB on an M5 at 153GB/s.

If you run local language models. Bandwidth decides everything here: token generation is bound by how fast weights are read out of memory. The gap between 307 and 614GB/s on M5 Pro and M5 Max shows up directly in response speed. This is the one scenario where I think M5 Max earns its price.

If you need iOS builds. There is no choice to make: Xcode runs on macOS only. The rest of the comparison stops mattering, and the question narrows to which Mac.

If it is backend and web. Almost anything in the table will do. Here I would look at cooling and the keyboard rather than the chip: eight hours in an editor is felt more sharply than ten percent in a benchmark.

On cooling specifically. Passive cooling is silent and pleasant right up until the load becomes sustained. After that the system drops clocks, and the peak figures from reviews stop describing your work. If you compute for hours, look for active cooling even at the cost of noise and weight.

The numbers that did not hold up

I checked the specifications against primary sources, and several widely repeated claims failed. Worth knowing what to watch for.

“Super cores arrived in M5.” They did not. Apple uses that term for M5 Pro and M5 Max and their 18-core CPU. The base M5 is described differently, and has no super cores.

“M5 Pro supports 128GB.” It does not. M5 Pro caps at 64GB and 307GB/s. The 128GB and 614GB/s figures belong to M5 Max. This one is expensive to get wrong: a table like that will sell you the wrong machine.

“9 to 15 percent faster single-threaded.” There is no official single-thread figure for the generation. Apple states up to 15 percent multithreaded and nothing else.

“Ryzen AI 9 does around 120GB/s.” Peak bandwidth is 89.6GB/s. The gap to Apple is wider than usually reported.

“Panther Lake is roughly 135GB/s.” No single figure exists. It depends on the SKU and varies by roughly a factor of two.

The general rule I took away: check specifications against vendor documentation, not against reviews. Reviews copy each other, and one error propagates across dozens of articles in a week.

Where I am leaning

For a long time the MECHREVO on a Ryzen AI 9 led my list: it offers the most memory per unit of money, and memory is my bottleneck. The logic was direct, take the capacity and live with the narrow bus.

I am now leaning towards Apple, and performance is not what decides it. iOS builds decide it. The mobile side of my products runs into the fact that Xcode exists only on macOS, and no amount of memory on Windows gets around that wall. The moment iOS appears on the task list, the platform comparison ends before it starts.

Inside the Apple line, choosing between an Air on M5 and a Pro on M5 Pro comes down to cooling and memory. The Air is quieter and lighter, but passive cooling will throttle on long runs, and a 32GB ceiling repeats the exact mistake I am replacing this machine over. The Pro removes both limits and costs considerably more.

I have not made the final call. But my criterion is now single and drawn from experience: do not buy a machine whose ceiling is visible on the day you buy it. Last time I took 16GB of soldered memory assuming it would last. It lasted eleven products and about three years. I am not walking into that twice.

What the table does not decide

Three years ago I chose on specifications and got none of them wrong. What I got wrong was different: I did not account for the work growing while the memory stayed fixed. Specifications describe a machine on the day you buy it, and you will be working on it two years later.

So the question I now ask first is not “how many cores” but “what here cannot be changed afterwards”. Soldered memory, a configuration ceiling, passive cooling: those are decisions made on your behalf, permanently. Clock speed and storage you can live with. A ceiling you cannot.

All specifications checked against Apple, AMD, Intel and Qualcomm documentation as of 20 August 2026. I have deliberately left prices out: they change faster than anything else in this piece.