Upstream Ollama accelerated on Intel Arc / Xe GPUs (iGPU + dGPU) via llama.cpp SYCL
592
The current upstream Ollama, accelerated on Intel GPUs through the llama.cpp SYCL backend (oneAPI / Level Zero).
Normal Ollama (ollama pull, ollama run, the API, Open WebUI), but with a SYCL GPU backend for
Intel Arc / Xe iGPUs (Meteor Lake, Arrow Lake, Lunar Lake) and dedicated Arc cards (A-series, B-series).
Rebuilt automatically for every new Ollama release.
Source, docs and issues: https://github.com/Dosperado74/ollarc-intel
This page is generated from
docs/dockerhub.md. The same images are also published asghcr.io/dosperado74/ollarc-intel.
ASUS NUC 14 Pro+, Core Ultra 9 285H (Arc 140T iGPU), tag latest-arl-h, 100 % GPU:
| Model | Prefill | Decode |
|---|---|---|
llama3.2:3b (Q4_K_M) | 1249 tok/s | 26.1 tok/s |
gpt-oss:20b (MXFP4) | 288 – 1106 tok/s | 14.0 tok/s |
gemma4:e4b | 361 tok/s | 16.1 tok/s |
gemma-4-26B-A4B-it-qat-q4_0 (MoE) | 184 tok/s | 14.6 tok/s |
More numbers and how they were measured: README.
The host needs a recent kernel (Arrow Lake ≥ 6.11) and the Intel GPU firmware
(Debian: firmware-intel-graphics; Ubuntu, Fedora, Proxmox VE: already included).
ls /dev/dri must show renderD128. Details: host setup.
docker run -d --name ollama --restart unless-stopped \
--device /dev/dri \
--group-add "$(stat -c %g /dev/dri/renderD128)" \
-v ollama:/root/.ollama -p 11434:11434 \
dosperado74/ollarc-intel:latest
docker exec ollama ollama run llama3.2:3b
docker exec ollama ollama ps # -> PROCESSOR 100% GPU
With Open WebUI: use the docker-compose.yml from the repository. Proxmox VE LXC: Proxmox guide.
| Tag | Kernels | For |
|---|---|---|
latest | JIT (SPIR-V) | recommended, every supported Intel GPU |
latest-arl-h | AOT arl-h | Arrow Lake-H iGPU (Core Ultra 200H, Arc 140T / 130T) |
latest-dg2 | AOT acm-g10,acm-g11 | dedicated Arc A-series (A770, A750, A580, A380, A310) |
latest-bmg | AOT bmg-g21 | dedicated Arc B-series (B580, B570, Arc Pro B50 / B60) |
Every tag also exists per Ollama version, e.g. 0.40.15, 0.40.15-arl-h.
AOT images only shorten the first model load; inference speed is the same as JIT.
MIT license. Not affiliated with Ollama or Intel.
Content type
Image
Digest
sha256:a61ee54eb…
Size
2.3 GB
Last updated
6 days ago
docker pull dosperado74/ollarc-intel