Discussion

PCWorld’s RTX Spark test shows why 128GB of unified memory can matter

In Mission Control

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#11239

PCWorld says an RTX Spark laptop completed a demanding run with OpenAI’s 120-billion-parameter gpt-oss model faster than a desktop with an RTX 5090. The result is a striking demonstration of what a large pool of GPU-accessible memory can do, not proof that the laptop is faster at AI workloads in general.

NVIDIA Watch analysis

What happened

In a test described in its 4 November review, PCWorld ran the same research workload on an Asus ProArt P16 with 128GB of unified memory and a desktop with an RTX 5090. The laptop completed 472,084 tokens in 33 minutes and 13 seconds; the desktop took 52 minutes and 51 seconds.

PCWorld says the laptop could make roughly 96GB of its memory available to the GPU, while the desktop’s RTX 5090 has 32GB of dedicated VRAM. The review says the RTX 5090 is faster when a workload fits in that faster memory. Here, the larger model and memory-hungry workload changed the contest.

Why it matters

This is a useful example of capacity beating raw GPU speed when a local model needs more memory than a conventional graphics card can supply directly. For people running large models on their own machines, that could mean workloads that would otherwise require a different system or cloud inference become possible locally.

There are substantial trade-offs. PCWorld’s tested laptop cost $7,000, and the review says its fans ran loudly while the AI workload used 46% of the battery in 33 minutes. It also reports software compatibility wrinkles, including needing a beta version of Adobe Premiere for proper GPU support.

Our read

The most interesting finding is not that a laptop has beaten an RTX 5090. It is that the laptop’s unusually large unified-memory pool helped on this particular job, while the desktop retained the advantage on workloads that fit its dedicated VRAM. That is a more useful buying insight than a leaderboard headline, if a somewhat less tidy one.

PCWorld’s result makes RTX Spark worth attention for people who regularly run large local models. It does not make a $7,000 laptop the sensible choice for everyone, or establish a general performance win. Check that your own models and software benefit from the extra memory before treating this as an upgrade prescription.

What to watch

  • Whether other reviewers reproduce the result with the same model and workload.
  • How RTX Spark performs across different local models and tasks that fit within a desktop GPU’s VRAM.
  • Whether software support and battery performance improve as the platform matures.

Discussion spark: For local AI, would you pay a premium for enough memory to run larger models on a laptop, or is cloud access better value until the software and battery trade-offs improve?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.