Intel Arc Pro B70 Open-Source Linux Performance Against NVIDIA RTX & AMD Radeon AI PRO

Written by Michael Larabel in Graphics Cards on 15 April 2026 at 01:00 PM EDT. Page 2 of 7. 8 Comments.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

First up was Llama.cpp with the respective Intel SYCL / AMD ROCm / NVIDIA CUDA back-ends. When running DeepSeek R1 Llama 8B across the tested graphics cards with the prompt processing 512 benchmark, the Arc Pro B70 was around the same performance as the prior-generation Radeon PRO W7900.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

But where the Arc Pro B70 becomes much more enticing is the sub-$1000 price point at the moment. With the Arc Pro B70 32GB able to be bought for $949 USD in the US, it does present a better bargain than the prior-generation NVIDIA RTX Ada graphics cards here and only the Radeon AI PRO R9700 was delivering better value.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 1024. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: DeepSeek-R1-Distill-Llama-8B-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

With the competitive pricing of the Arc Pro B70, it's quite positioned well for those seeking good value on their hardware. Plus, like with the AMD graphics cards, is backed by a fully open-source Linux graphics/compute driver stack.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

For Llama 3.1 Tulu 3 8B, the prompt processing performance of thr Arc Pro B70 was similar to the Radeon PRO W7900 with Llama.cpp while delivering better value than the tested NVIDIA RTX graphics cards.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: SYCL / ROCm / CUDA, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

The GPT-OSS 20B performance of the Arc Pro B70 was coming up short under Llama.cpp. This may be due to Llama.cpp issues or lack of optimizations on the Intel side, but will be interesting to see how the vLLM Scaler performance compares. There I am currently working through some Docker issues and setting up a good workflow for that comparison with vLLM albeit made messier by Intel not yet having as nice upstream support as AMD ROCm and NVIDIA CUDA in the project.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Llama-3.1-Tulu-3-8B-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

Testing these graphics cards with Llama.cpp's Vulkan back-end showed more room for improvement with the Intel ANV Vulkan Linux driver.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 1024. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Mistral-7B-Instruct-v0.3-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

Particularly for pp1024 and pp2048, the Intel ANV Vulkan performance was dropping heavily.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: gpt-oss-20b-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3.5-9B-Q8_0, Test: Prompt Processing 512. RTX 6000 Ada Gen was the fastest.

👁 Llama.cpp benchmark with settings of Backend: Vulkan, Model: Qwen3.5-9B-Q8_0, Test: Prompt Processing 2048. RTX 6000 Ada Gen was the fastest.

8 Comments - Next Page

Page: 1 2 3 4 5 6 7 Next Page

URL: https://www.phoronix.com/review/intel-arc-pro-b70/2

⇱ Intel Arc Pro B70 Open-Source Linux Performance Against NVIDIA RTX & AMD Radeon AI PRO Review - Phoronix

Intel Arc Pro B70 Open-Source Linux Performance Against NVIDIA RTX & AMD Radeon AI PRO