On stage, Jensen Huang holds up a small machine. According to NVIDIA's live blog of the keynote, Huang said that in 2016 he delivered 1 petaflop and 128GB to a small company called OpenAI 1. Holding up the machine, he went on: "This is 1 petaflop and 128GB" 1, adding, "The only difference, of course, is that I can hold this like this, and DGX-1 took two people to carry upstairs." 1
According to NVIDIA's live blog, late in the GTC Berlin keynote Huang turned to RTX Spark laptops, which had launched the week before 2. He compared their 1 petaflop of performance and 128GB of memory with the DGX-1 he delivered to OpenAI in 2016, and made the difference in size the punchline 3. That is NVIDIA's own way of comparing; it is not a claim that the old server and the laptop perform the same.
This piece reads several voices separately.
- The makers' announcements: NVIDIA's keynote and blog, Adobe's press release
- Introductory posts on X: a poster's own summary, treated as the poster's claim
- Outlets: numbers PC Watch measured, for the machine and settings it used, and ITmedia PC USER's outlook
The aim is to set "what was promised" next to "what was measured" so you can compare them yourself.
A demo of an agent helping train a robot duck
According to NVIDIA's blog, a video shown in the keynote had an agent helping train Hugging Face's Microduck. It set up a course and simulated the robot in Isaac Sim, sent training jobs to the cloud, then brought the learned policy back to RTX Spark for evaluation before deploying it to the physical robot 4.
Four Microducks joined Huang onstage, sitting down, standing up and even taking a nap on cue 5. He explained that they were created on RTX Spark, trained in the cloud and learned their skills in simulation 5.
In a separate post, NVIDIA's blog says that during the keynote Huang showed how AI agents on RTX Spark can be used to train the Microduck robot using NVIDIA Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B 6.
On X, a Turkish-language post describes a Hermes Agent demo from the keynote. The poster writes (in Turkish) that Huang closed the keynote with a Hermes Agent demo 7. According to the poster, a single prompt had Hermes Agent installed on an RTX Spark laptop via OpenShell, with the instruction "pick the best open model that fits this machine, …" (translated from the Turkish original; the post breaks off there) 8.
Jensen Huang, GTC Berlin keynote’unu bir Hermes Agent demosuyla kapatmış.
— Erhan Meydan (@erhanmeydan) October 21, 2026
"Bir agent yapmanın yolu, onu bir agent’a yaptırmak" dedi :))
Demoda olanları;
→ Tek bir prompt ile Hermes Agent’ı RTX Spark dizüstüne OpenShell ile kurduruyor, “bu makineye sığan en iyi açık modeli seç,…
In NVIDIA's live blog, the talk actually ended with a Eurovision-style musical send-off that recapped its themes 9. Read "closed with a Hermes Agent demo" as the poster's summary only.
Adobe's announcement: Photoshop and Premiere for RTX Spark
Adobe announced that Photoshop, Premiere and Substance 3D, optimized for RTX Spark, speed up AI processing, editing, color and effects work by up to 2x 10. The "up to 2x" covers the three apps together, and the release does not say what the figure is measured against 10. The Japanese version of the release is an abridged translation of the press release issued by Adobe's US headquarters on October 15, US time 11.
On Premiere, Adobe says some of the most demanding post-production workflows, including playback, rendering, effects, color and AI tools, are accelerated with RTX and optimized for RTX Spark. It says this speeds up popular features such as "Enhance Speech" and "Object Mask" (translated from the Japanese release) 12. Adobe writes that Photoshop was redesigned around GPU-accelerated compositing, with faster, more responsive tools such as Remove and effects applied directly on the canvas 13. NVIDIA also says it is working with Adobe to redesign Premiere and Photoshop for RTX Spark 14.
On X, a Japanese-language post introduces an NVIDIA official video. The poster writes (in Japanese) that NVIDIA put 2 RTX Spark laptops of the same kind side by side in the video to demonstrate the performance gap between the standard and the accelerated versions of Premiere (the post text is cut off) 15.
NVIDIAは公式動画「NVIDIA RTX SparkでAdobe Premiereの編集を速く」において、同じRTX SparkのノートPC2台を並べ、Adobe Premiere(動画編集ソフト)の通常版と高速化版の性能差を実演した。
— 清水勇介｜Yusuke Shimizu (@ShimizuYusuke) October 22, 2026
・この動画では、Adobe… pic.twitter.com/LzRMmJtcHo
Separate from the announcements: what outlets measured and expect
From here on these are outlet measurements, not announcements. PC Watch tested Photoshop Beta and Premiere Beta on the 64GB ASUS ProArt P16 (H7607) 16 17.
- Photoshop Beta: PC Watch ran the Remove tool with generative AI on and on-device processing, timing it by hand and excluding the time spent marking targets with the brush [[F253]]. Removing 4 objects at once from an indoor photo (a hand sculpture, a black sphere, and plants on the left and right) took 11.89 seconds [[F252]]
- Premiere Beta: the Object Mask process on about 29 seconds of footage with a car in it took 30.79 seconds, including initialization [[F245]]
Both are Beta-version results from this one machine. PC Watch did not measure whatever Adobe's claim is compared against, so these numbers cannot be read as support for Adobe's "up to 2x" 10. The announced figure and the measured figures answer different questions.
In its ASUS ProArt P14 (H7407) review, ITmedia PC USER writes that improvements to the environment, such as better OS memory management and optimization of Photoshop and Premiere Pro, are planned 18. The same review adds that the trial period was short and the work had reached the stage of using the machine and running benchmarks 19.
Big models: NVIDIA's description and the measurements
NVIDIA says RTX Spark can run models up to 120B parameters with 1M context 20. According to NVIDIA's porting guide, memory is up to 128 GB, shared by the CPU and GPU, with a bandwidth of 300 GB/s 21.
So what did the outlets measure? PC Watch ran a 27B LLM (Qwen3.8-27B, Q4_K_M, about 16.45GB) on the 64GB machine 22. It adds that even with weights of about 16.45GB, a model uses other memory at run time, so you should not decide the capacity you need from the file size alone 23. At an input of 65,536 tokens, the first output took about 102 seconds, and PC Watch notes that "you can hand it a long text" and "it answers right away" are not the same thing 24. The same review also says that fitting a large model and running it fast are different matters 25.
PC Watch's review of the 128GB ProArt P14 writes that it can load a model of over 90GB, even the 3-bit quantized Unsloth version, in full 26.
If you want to try it on your own desk
The "RTX Spark" in all of the above is limited to what was announced and demonstrated. Adobe has announced RTX Spark optimizations for Premiere 12, but its text says they cover some of the most demanding workflows 12. Before buying, check the official supported environments for the tools you want to run.
PC Watch's verdict on the 64GB machine is that 64GB is worth considering if you switch between a single quantized LLM or generation workflow and it all fits, and that 128GB is the better choice if the models or the number of simultaneous jobs will grow 27.
Ad: the “Search on Amazon” buttons are affiliate links, and we may earn a commission. They open a search, not a specific listing.
Ad: the “Search on Amazon” buttons are affiliate links, and we may earn a commission. They open a search, not a specific listing.
You can also look for RTX Spark laptops on Amazon.
Big models and generated files also need somewhere to live.