Browser Built-in AI Takes Too Long To Give Results

94 views
Skip to first unread message

C K.K

unread,
Oct 14, 2025, 2:14:42 PMOct 14
to Chrome Built-in AI Early Preview Program Discussions
Hi Google Chrome Team,

Celestino here. Integrated Prompt API and Summarizer API, but the time to output takes too long, specifically the Summarizer API takes a total of 133 seconds when I ask it to summarize a product page for the extension's user on Amazon for example. 
SummarizerAIAPIUse.png


I think this is something you can work on, unless there is a way I can optimize this further on Chrome settings? I need it to take very few seconds, like the native experience I get when using Gemini Web App or ChatGPT

Kindly advise.

Please see screenshotsSummarizerAPITiming.png

Kind Regards,
Celestino

C K.K

unread,
Oct 15, 2025, 3:08:02 AMOct 15
to Chrome Built-in AI Early Preview Program Discussions, C K.K

It takes now up to 4 minutes to successfully call the Prompt API 

Connie Leung

unread,
Oct 15, 2025, 3:10:30 AMOct 15
to C K.K, Chrome Built-in AI Early Preview Program Discussions
Will streaming response improve the user experience?

--
You received this message because you are subscribed to the Google Groups "Chrome Built-in AI Early Preview Program Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chrome-ai-dev-previe...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/6da675fd-d4b7-4ca2-8881-98f6886219afn%40chromium.org.

Thomas Steiner

unread,
Oct 15, 2025, 3:26:41 AMOct 15
to Connie Leung, C K.K, Chrome Built-in AI Early Preview Program Discussions
Hi Celestino,

It's a known challenge that the API has a bit of a "cold start" problem where the first prompt would take a couple of seconds until the first token is emitted, but four minutes sounds excessive. Does this reproduce on another device? If you're using the `summarize()` method, can you try the `summarizeStreaming()` method (even if you end up just concatenating the chunks and work with the final result)? I'm interested in hearing your results for when the first token is emitted. Happy to take a look at the code as well, feel free to share it in private if it can't be shared publicly.  

Cheers,
Tom

Thomas Steiner, PhD—Developer Relations Engineer (blog.tomayac.comtoot.cafe/@tomayac)

Google Spain, S.L.U.
Torre Picasso, Pl. Pablo Ruiz Picasso, 1, Tetuán, 28020 Madrid, Spain

CIF: B63272603
Inscrita en el Registro Mercantil de Madrid, sección 8, Hoja M­-435397 Tomo 24227 Folio 25

----- BEGIN PGP SIGNATURE -----
Version: GnuPG v2.4.8 (GNU/Linux)

iFy0uwAntT0bE3xtRa5AfeCheCkthAtTh3reSabiGbl0ck
0fjumBl3DCharaCTersAttH3b0ttom.xKcd.cOm/1181.
----- END PGP SIGNATURE -----

Kenji Baheux

unread,
Oct 27, 2025, 2:03:37 AM (4 days ago) Oct 27
to C K.K, Connie Leung, Thomas Steiner, Chrome Built-in AI Early Preview Program Discussions
Hi Celestino,

Sorry to hear about such high latency.

Would you mind sharing details about your device (OS, CPU, GPU, Memory RAM and/or VRAM, SSD or HDD, etc) so we can investigate this further?
Also, a GPU report downloaded from chrome://gpu could be helpful. Probably best to send this directly to me instead of the group.

Thanks in advance!



--
Kenji BAHEUX (my how-to)
Product Manager - Chrome
Google Japan

C K.K

unread,
Oct 27, 2025, 1:32:52 PM (4 days ago) Oct 27
to Chrome Built-in AI Early Preview Program Discussions, Connie Leung, Chrome Built-in AI Early Preview Program Discussions, C K.K
Hi Connie, 

I have tried adopting the streaming experience and still no improvements

C K.K

unread,
Oct 27, 2025, 1:33:14 PM (4 days ago) Oct 27
to Chrome Built-in AI Early Preview Program Discussions, Thomas Steiner, C K.K, Chrome Built-in AI Early Preview Program Discussions, Connie Leung
Hi Thomas,

Still lots of problems, sharing code in an hour or less

C K.K

unread,
Oct 27, 2025, 1:36:46 PM (4 days ago) Oct 27
to Chrome Built-in AI Early Preview Program Discussions, Kenji Baheux, Connie Leung, Thomas Steiner, Chrome Built-in AI Early Preview Program Discussions
Attached is the GPU report, and system information(dxdiag file)

about-gpu-2025-10-27T17-33-29-500Z.txt
DxDiag.txt

Kenji Baheux

unread,
Oct 27, 2025, 7:02:56 PM (3 days ago) Oct 27
to C K.K, Chrome Built-in AI Early Preview Program Discussions
From a quick look, there seems to be a problem with your GPU AMD Vega 10 (error code 43) which results in having it disabled in Chrome.
An online search suggests that it might be a driver problem. Can you uninstall and reinstall your AMD drivers?

On Tue, Oct 28, 2025 at 2:36 AM C K.K <kcelestin...@gmail.com> wrote:
Attached is the GPU report, and system information(dxdiag file)



C K.K

unread,
Oct 27, 2025, 7:04:49 PM (3 days ago) Oct 27
to Kenji Baheux, Chrome Built-in AI Early Preview Program Discussions
My display driver has a problem and HP nor AMD has updated the drivers. So do you mean the display driver affects Chrome AI API compatibility?

C K.K

unread,
Oct 29, 2025, 4:06:19 PM (2 days ago) Oct 29
to Chrome Built-in AI Early Preview Program Discussions, C K.K, Chrome Built-in AI Early Preview Program Discussions, Kenji Baheux
Hi Kenji, Tom & Connie,

I'm going to have to submit it anyways. even when I install and uninstall the drivers, no luck, I wish we can decouple the AI APIs and how they work from many of the resources and operations at the OS level like drivers.

C K.K

unread,
Oct 29, 2025, 4:06:43 PM (2 days ago) Oct 29
to Chrome Built-in AI Early Preview Program Discussions, C K.K, Chrome Built-in AI Early Preview Program Discussions, Kenji Baheux
I'll still be building after the hackathon :) 

Kenji Baheux

unread,
2:12 AM (3 hours ago) 2:12 AM
to C K.K, Chrome Built-in AI Early Preview Program Discussions
In general, driver issues can make the GPU unavailable or unstable and therefore not being used in Chrome (AI or not).
In the logs, there is a fallback "GPU" in the form of Microsoft Basic Render Driver but it's blocked because it crashes too often or is too slow.

That said, if you have at least 16GB of RAM then you should automatically fallback to the CPU inferencing / model.
Can you try Chrome canary?

Reply all
Reply to author
Forward
0 new messages