Choosing a desktop replacement laptop in 2025
Summary
A desktop replacement laptop is real. Although you should move your focus from mass market slip and light laptops and look on game laptops, workstations and googled 'desktop replacement laptop'.
The most underestimated thing is CPU throttling. Under the critical temperature (~100C) CPU slow downs its work automatically. It means the numbers CPU has in its description are nominal, the actual performance is unclear.
It seems like a good strategy to look for processor with maximized base frequency and the market has several good examples, but in general CPU manufactures goes by a different way -- create a CPU with a really vast gap between base and boost performance and let PC\laptop manufacturers choose desired mode by playing with cooling system. It was said the edge innovation nowadays goes not in processor itself, but in the cooling system.
Anyway 55 Watts is an optimum for old styled laptops. Further grows of power consumption increase heating which leads to throttling -- slow down performance. It seems like more than 8 core CPU is not worth to install into such laptops.
However, innovative cooling system helps to reach sustained performance even in much higher power budget. I haven't checked it yet by my own, but laptop I just ordered has such configuration. Will see how it would work.
The second bottleneck is disk and network speed. SSD really made a difference compare to HDD. Evolution of PCI lines doesn't make such difference, but each next version doubles speed of transmitting data -- very important point to pay attention to. PCI 5 exists on the market and PCI 6 is going to be introduced.
There is a rising trend of professionals who looks for a laptop as a replacement of their stationary PC. Manufacturers need time to adopt, but even now good examples of such laptops exist on the market -- Lenovo Legion 7 pro and MSI Crosshair are ones of them.
Finally, you have to decide how much money you are going to spend on a new machine and how often you will do this. Technology evolves and software requirements growth as well. In means performance capacity with time would not be enough for a new tasks and next upgrade would be required. Investing more finance now to buy a better machine prolonges its life, but investing smaller budget could be cost effective: not overpaying for cutting edge solutions and benefit from new generation of technology quicker due shorter periods of upgrades.
Personally, I am not satisfied with what market offers right now for my aims. Yes, CPU doubles its raw power compare to my current laptop, but this power could NOT be released due temperature limits of laptops. My research results tend to conclude if you can wait it would be good to postpone upgrade for two years.
| Period | Upgrade? | Why |
|---|---|---|
| 2025–2026 | ❌ | Evolution without sustained growth |
| 2027 | ❌ | Still OEM limitations |
| 2028–2029 | ✅ YES | Increase in base clocks + RAM + I/O |
| 2030+ | ✅ | New class of devices |
p.s.
There is a seasonal periodic dynamic in prices for computers:
- highest prices are a few months before New Year
- lowest prices in March
- in June prices could be even lower, but the variety of options is
narrowed down until upcoming September
Motivation
After 10 years of using my current laptop, it became clear that an upgrade is necessary.
Modern processors offer performance many times higher than my current machine, RAM limits
are reached (16 GB is now insufficient for engineering and ML tasks), and some frameworks
no longer support my GPU.
The goal is to select a laptop capable of handling sustained multi-threaded workloads,
ML workloads, and modern storage requirements while remaining energy-efficient and avoiding
throttling under heavy use. This study explores the components, trade-offs, and best practices
for selecting such a laptop in 2025.
Processor
For engineering workloads, effective multi-threading is key, but
single thread performance is important either. However, laptops
with more than 8–16 cores often do not provide linear performance
gains due to thermal limitations and throttling. Key considerations:
- Maximum practical TDP: 55 W — higher TDP rarely yields sustainable gains.
- Base frequency is more important than turbo: base frequency reflects stable performance under prolonged load.
- Ryzen 7/9 HS (8 cores, 3.8–4.0 GHz base) is excellent for engineering tasks. These CPUs often outperform
competitors in sustained performance and energy efficiency.
- Intel HX / Ultra HX chips offer more cores and high turbo, but lower base frequency and higher TDP can
result in throttling under long-term loads.
However in general CPU manufactures goes by a different way -- create a CPU
with a really vast gap between base and boost performance and let
PC\laptop manufacturers choose desired mode by playing with
cooling system. It was said the edge innovation nowadays goes not
in processor itself, but in the cooling system. (still there is a limit what
you can do with cooling system)
Storage
NVMe SSD, M.2 slot, x4 lanes. PCIe Gen 4/5 preferred; Gen 6 desirable but rare.
| PCIe generation | Data rate (GT/s per lane) | Max throughput (x4) | Typical NVMe SSD speeds (sequential) | Year/status |
|---|---|---|---|---|
| Gen 1 | 2.5 | ~1 GB/s (~8 Gb/s) | ~400–500 MB/s | Early 2000s |
| Gen 2 | 5 | ~2 GB/s | ~1 000 MB/s | Later 2000s |
| Gen 3 | 8 | ~4 GB/s | ~3 000–3 500 MB/s | Widespread |
| Gen 4 | 16 | ~8 GB/s | ~6 000–7 000 MB/s | Modern mainstream |
| Gen 5 | 32 | ~16 GB/s | ~10 000–14 000 MB/s+ | Latest PCs |
| Gen 6 | 64 (PAM4) | ~32 GB/s | ~20 000–28 000 MB/s (future SSD) | Emerging ([atpinc.com](https://www.atpinc.com/de/blog/PCIe-vs-NVMe-are-they…ource=chatgpt.com)) |
Further research shown migration from PCI 4 to PCI 5 and further gives boost in performance for only specific set of tasks -- it is good for server and some narrow common user scenario on PC, but for the rest it would not bring much gains. The issue is a latency in access to the disk, it becomes a modern bottleneck. Solution for this would be in Storage Class Memory / CXL / persistent memory / byte-addressable NVM.
One more important point is heat budget. Current generation of SSD based on PCI 5 nearly doubles&triples power consumption and produced heat compare to SSD PCI 4. Why is it important? Because heat budget is a very limited on laptops, nowadays heat is a major factor which limited performance of your laptop and blocking to unleash its full power. That's why it is important to minimize heat produced by each component of the computer with remain the performance capacity.
RAM
Dual-channel or quad-channel configuration is highly recommended to maximize memory bandwidth. 32–64 GB is ideal for current engineering and ML workloads.
By reports extending RAM is not a trivial task due compatibility issues. You might need several try & run iterations for find one which fits your machine. Anyway you should expect increasing RAM would lead to decrease of working frequency.
If you plan to use you PC\laptop more than 4 years, consider machines which motherboard and CPU supports to double current amount of RAM.
GPU
Discrete GPU is desirable for ML tasks. Performance is important, but consider thermal and power limits.
By reports 8Gb GPU is not enough for neural networks which are used in Photoshop and Lightroom. It means if you plan to train even moderate ML models on your machine, you will need at least 12Gb GPU.
All modern frameworks support only NVidia. For ML tasks there is no other choice.
Cooling
Sustained performance is critical. Efficient cooling prevents CPU/GPU throttling under long workloads.
Vapor chamber, several fans, liquid metal are all cutting edge tech in cooling system design. Lenovo Legion 7 pro series seems to be a leader in this.
Networking
Wi-Fi 6/6E for wireless; 2.5–10 GbE Ethernet for wired connections.
BIOS & Power Settings
- Enable XMP/DOCP to run RAM at rated frequency.
- Set Windows High Performance profile.
- If available, tune PL1/PL2 and Turbo settings in BIOS.
- Monitor temperatures and optimize cooling to prevent throttling.
It is really nice to have an option to undervolt CPU cores or setup temperature limits in BIOS. In case of Lenovo, on Windows they have software which automatically tune CPU and fans to keep sustainable performance, but on Linux machine there is not such software -- all of such things should be done manually.