Does the Surface Laptop Ultra Need 48GB of RAM?
機械翻訳 / Machine-translated
A
AI Friends
8時間前・16分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Are you wondering, "I want to run local AI, but how many GB of RAM do I actually need?" Using Microsoft's new laptop, the Surface Laptop Ultra, as an example, we'll break down how to choose the right memory capacity and how it compares to the competition.
At an event held on October 7, 2026 (North American time), Microsoft announced the Surface Laptop Ultra, powered by NVIDIA's new chip, the "RTX Spark N1X."
The launch date is October 16, 2026, and pre-orders are already open.
The display is a 15-inch mini-LED panel (a bright LCD with densely packed tiny LEDs), with a peak brightness of 2,000 nits. The laptop is approximately 19mm thick and weighs around 2kg.
Prices (tax included) are as follows:
The top-end 128GB configuration had already sold out by 5:00 AM on October 8, the day after the announcement — a testament to just how much attention this device has attracted.
According to Petri's coverage, the RTX Spark integrates an Arm-based CPU with a Blackwell-generation GPU into a single chip. Its AI performance is rated at up to 1 petaflop (one quadrillion calculations per second).
A key feature is unified memory — a setup where the CPU and GPU share the same pool of memory, with configurations available up to 128GB.
In a typical PC, the dedicated GPU memory (VRAM) available for AI is around 8–16GB. If a model is larger than that, it simply won't run.
With unified memory, the bulk of the system RAM can be allocated to AI workloads. This is what makes it possible to run large AI models that were previously impractical on a laptop.
ASCII.jp's hands-on article states that to take full advantage of the RTX Spark, a minimum of 48GB is needed — with 64GB or 128GB being ideal.
At the same time, it concludes that 32GB is "nowhere near enough for serious use."
The rough rule of thumb for memory capacity is "number of parameters (a measure of the AI model's size) × bit count." Using quantization — a technique that compresses a model to reduce its bit count — you can lower the memory requirements, but only so much.
Here are the examples ASCII.jp provided:
In other words, even with compression, large models still demand considerable memory. And on a PC, the OS and applications also consume RAM. The "48GB or more" figure, with room to spare, is easy to understand and justify.
The 18-core CPU models are limited to 24GB or 32GB of memory. ASCII.jp describes this as "underwhelming" for local AI use.
Since these capacities aren't much different from existing PCs, the article questions what the point of the RTX Spark even is at these configurations.
Imagine a salaried worker who buys the ¥513,480 24GB model. Small AI models for summarizing work documents would run fine. But when they want to try one of the popular larger models, they'd hit the memory ceiling and have to give up.
Stepping up to 48GB or more pushes the cost to around ¥760,000 or higher. This device simply isn't suited for the "budget entry into local AI" approach.
Even with enough capacity, that doesn't necessarily mean it'll run fast. The speed at which AI generates text depends on the memory's bandwidth — the amount of data that can be read and written per second.
The bandwidth NVIDIA has publicly stated is 300 GB/s via a 256-bit bus. ASCII.jp notes this is slightly lower than the GeForce RTX 5050's 320 GB/s.
From this figure, ASCII.jp interprets the design philosophy as prioritizing the size of models that can be loaded over raw processing speed.
Think of it this way: the warehouse is massive, but the road leading to it is only standard width. You can store large models, but that doesn't mean text will stream out smoothly.
Other laptops feature unified memory architectures as well. Here's a look at the main competitors.
According to EveZone's comparison article, the M5 Max supports up to 128GB with a maximum bandwidth of 614 GB/s — roughly twice that of the RTX Spark.
The same article reports that running a 70B-class model (Q4_K_M quantization) on a 128GB M5 Max yields around 12–15 tokens per second.
This is another existing option that ASCII.jp mentions. According to AMD's technical blog, it supports up to 128GB of memory, with up to 64GB available to the GPU by default.
Benchmark results from the same blog showed approximately 42 tokens per second for a 35B-class MoE model (a type that only activates the parts it needs), and approximately 8.6 tokens per second for a 122B-class model.
To put it simply: if speed is your priority, Mac wins; if you want flexibility in options and pricing, Ryzen is worth considering; and if you want NVIDIA's AI ecosystem on Windows, Surface is your pick.
That said, the numbers cited here were measured differently by each source. ASCII.jp also notes that battery-powered performance will be verified in future hands-on testing, so it's safer to wait for those results before making a final judgment on real-world speed.
Yen-denominated pricing has been announced in Japan, and the device goes on sale October 16 — no need for personal imports.
There's also a tailwind on the software side: according to ASCII.jp, the new Nemotron model becomes available on October 15.
Many Japanese companies have departments that are cautious about sending data outside their walls. Teams handling design documents or customer records, for example, would naturally prefer running AI locally on their own machines rather than using the cloud.
For those use cases, a configuration loaded with RAM at around ¥1,000,000 becomes a very real consideration.
For individuals, price is the biggest barrier. It's probably safer to start by experimenting with smaller models on your current PC, and only consider upgrading once you've outgrown what you have.
October 16, 2026. Pre-orders are already available.
ASCII.jp recommends a minimum of 48GB. Ideally, 64GB or 128GB. It states that 32GB is insufficient for serious use.
Small models will run. However, the 18-core version tops out at 24GB or 32GB of RAM, making it unsuitable for larger models.
The key difference is memory bandwidth. The RTX Spark offers around 300 GB/s, while the M5 Max goes up to 614 GB/s. If speed is a priority, the Mac has the advantage.
It was sold out as of 5:00 AM on October 8. Restocking availability has not been confirmed.
As a next step, start by researching the size of the AI model you want to run, determine the memory you'll need, and then choose your device accordingly.
This article is a cross-post from AI Friends.
