Skip to content

[SYCL][Intel] Low performance on MoE models - SYCL is slower than VULKAN (A770) #19918

Closed as not planned
Closed as not planned
[SYCL][Intel] Low performance on MoE models - SYCL is slower than VULKAN (A770)#19918
Labels
stale
@savvadesogle

Description

@savvadesogle
savvadesogle
opened on Feb 26, 2026

Good day, Jianyu ❤️
@NeoZhangJianyu

I have problems with the performance of the A770 model with the SYCL backend.
PP: 600 vs 1100 t/s
TG: 10 vs 68 t/s

Models:
unsloth/gpt-oss-20b-GGUF https://huggingface.co/unsloth/gpt-oss-20b-GGUF
lmstudio-community/Qwen3.5-35B-A3B-GGUF https://huggingface.co/lmstudio-community/Qwen3.5-35B-A3B-GGUF

Image

SYCL lmstudio-community\Qwen3.5-35B-A3B-GGUF

Image Image

Vulkan lmstudio-community\Qwen3.5-35B-A3B-GGUF

Image Image

Tests with GPT-OSS 20B

Image 
C:\llm\llama-cpp\SYCL\b8157>llama-cli --version
load_backend: loaded RPC backend from C:\llm\llama-cpp\SYCL\b8157\ggml-rpc.dll
load_backend: loaded SYCL backend from C:\llm\llama-cpp\SYCL\b8157\ggml-sycl.dll
load_backend: loaded CPU backend from C:\llm\llama-cpp\SYCL\b8157\ggml-cpu-haswell.dll
version: 8157 (2943210c1)
built with Clang 19.1.5 for Windows x86_64

Metadata

Metadata

Assignees

No one assigned

    Labels

    stale

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions