LLM/Intel Pro Arc B70/2026-06-20

Setup ssh key for github.com

Host         github.com
IdentityFile ~/.ssh/xxx.pem

Download source

sudo apt install git cmake
sudo apt install libssl-dev openssl
mkdir -p ~/llama.cpp/src
git clone https://github.com/ggerganov/llama.cpp ~/llama.cpp/src

Build

Test CPU version

# build
cmake -B build -DGGML_NATIVE=ON
cmake --build build --config Release

# clean
cmake --build build --target clean

Test OpenVINO version

# setup apt source for OpenVINO
wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
sudo gpg --output /etc/apt/trusted.gpg.d/intel.gpg --dearmor GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
echo "deb https://apt.repos.intel.com/openvino ubuntu24 main" | sudo tee /etc/apt/sources.list.d/intel-openvino.list

# install OpenVINO and dependencies
sudo apt update
sudo apt install -y openvino
sudo apt install -y libtbb-dev opencl-c-headers ocl-icd-opencl-dev

# build
cmake -B build-openvino -DGGML_OPENVINO=ON -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib"
cmake --build build-openvino --config Release

# Install 
cmake --install build-openvino --prefix ~/llama.cpp/openvino

# clean
cmake --build build-openvino --target clean

ln -sf $(find /usr/lib/ -name "libopenvino_intel_gpu_plugin.so" 2>/dev/null | head -n 1) ~/llama.cpp/openvino/lib/
cd ~/llama.cpp/openvino/bin
./llama-cli --list-devices

It sucks, no GPU found.

Available devices:
  OPENVINO0: OpenVINO Runtime (95663 MiB, 95663 MiB free)

Test SYCL(oneAPI) version

Install dependencies

# install oneAPI
sudo apt install intel-oneapi-mkl-devel
sudo apt install intel-oneapi-compiler-dpcpp-cpp
source /opt/intel/oneapi/setvars.sh

Link shared library

:: initializing oneAPI environment ...
   bash: BASH_VERSION = 5.2.21(1)-release
   args: Using "$@" for setvars.sh arguments: --force
:: compiler -- latest
:: debugger -- latest
:: dev-utilities -- latest
:: dpl -- latest
:: mkl -- latest
:: tbb -- latest
:: tcm -- latest
:: umf -- latest
:: oneAPI environment initialized ::

Build

# build
cmake -B build-sycl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DGGML_SYCL=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib"
cmake --build build-sycl --config Release

# Install 
cmake --install build-sycl --prefix ~/llama.cpp/sycl

# clean
cmake --build build-sycl --target clean

# test GPU
cd ~/llama.cpp/openvino/bin
./llama-cli --list-devices

Got It!

Available devices:
  SYCL0: Intel(R) Arc(TM) Pro B70 Graphics (32656 MiB, 32128 MiB free)

Test Models

# test model
./llama-server -hf google/gemma-4-2B-it-qat-q4_0-gguf --host 0.0.0.0 --port 9004     # 90 t/s
./llama-server -hf google/gemma-4-E4B-it-qat-q4_0-gguf --host 0.0.0.0 --port 9004     # 90 t/s
./llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf --host 0.0.0.0 --port 9012     # 47 t/s
./llama-server -hf google/gemma-4-26B-A4B-it-qat-q4_0-gguf --host 0.0.0.0 --port 9026 # 52 t/s
./llama-server -hf google/gemma-4-31B-it-qat-q4_0-gguf --host 0.0.0.0 --port 9031     # 23 t/s