Feature Request: improve SYCL docs - performance comparison & fix inconsistency to project-level README.md

[kha84]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Remove from the project-level README.md the mentioning of Nvidia being supported by SYCL

Currently the top project-level README.md page says SYCL backend can be used for Nvidia:

Backend	Target devices
...	...
SYCL	Intel and Nvidia GPU
...	...

but from the recent news I read:

2026.02
Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. User can't build up the software for Nvidia & AMD GPU.

This is inconsistecy

Add some short performance comparison table for the reference

It would greatly help to newcomers to understand what to expect from performance standpoint, when SYCL is used as a backend, compared say to Vulcan. I'm not expecting a wide-range comparison, but just few typical models would help a lot, like this:

card	model	backend	pp512	tg128
Arc b580	Qwen3.5-9B-Q4_K_M.gguf	Vulcan	???	???
Arc b580	Qwen3.5-9B-Q4_K_M.gguf	OpenVino	???	???
Arc b580	Qwen3.5-9B-Q4_K_M.gguf	SYCL	???	???

Without this table, people will be running llama.cpp with using SYCL backend blindly, questioning if they're reaching the reference performance or not, wasting their time on troubleshooting issues they might not have. Having reference performance numbers will benefit community.

NB: I do not own the Intel card unfortunately, otherwise I would gladly help you to benchmark it

Motivation

With recent release of powerful Intel Panther Lake APUs, there will be more interest from community in various backends that are focused on support Intel iGPU / dGPU

Possible Implementation

No response

[kha84]

I saw some work is done with documentation #23025 and I left my comment there, raise a separate issue just to ensure it won't be lost

[arthw]

@kha84
Yes, the description of Nvidia should be removed from the guide.

To add the short performance comparison table, it's not good idea:

1. It's not polite to list the performance data of other backend in SYCL document.

We are not the expert of other backend.
Maybe the test condition can show the best performance of other backend.
It will bring unnecessary trouble.

2. We (SYCL) don't want to challenge to other backends, through we always be challenged by others' user in SYCL issue.

But it's good idea to show the performance data in SYCL guide as reference.
I know the SYCL users always sharing the local test performance data of LLMs on different Intel GPU.
We could collect them.
Maybe they are not real time data for latest code.

Let's consider the solution:

1. update the data in SYCL.md file.
   easy to read, but update it hard due to workflow of llama.cpp.

2. update the data in a file of another repo, add a link in SYCL.md.
   easy and quick to update without CI, user need to click the url to read the data.

How do you think?

Thank you!

[kha84]

I fully get the point of questionable politeness of side-by-side comparison to other backends. So you certainly can leave it up to the user to find other's figures himself to compare them. As for the rest - sounds great! Thanks for replying on this.

[kha84]

I have three bookmarked links where ppl are contributing their performance figures, and some single kind-hearted person carefully updates his initial post to include all the incoming flow. If you could do something like this - that will be awesome!

CUDA - #15013
Vulcan - #10879
ROCM - #15021

[arthw]

Yes, it's good idea.
We will add the discussion issue with the performance data of SYCL.

Thank you!

[arthw]

Add: #23313
I will update in SYCL.md later.

Thank you!

[arthw]

Update by PR: #23315

[mrigankad]

This appears to have been addressed by commit aabee04 (PR #23315), which:

• Removed unsupported GPUs from the SYCL line in README.md
• Added a performance reference chapter to docs/backend/SYCL.md

The issue can be closed.