Intel's $1,299 AI Card Runs AI at Half the Speed of a $6,500 NVIDIA Card

A tech enthusiast recently compared Intel's new Arc Pro B70 graphics card, which costs $1,299, to NVIDIA's high-end RTX 5090, which sells for over $6,500. The test involved running a large AI model called Qwen3.8-27B. Out of the box, the Intel card performed poorly, but after some software adjustments, it achieved about half the speed of the NVIDIA card for generating text and about one-third the speed for processing prompts. That means you can run AI tasks at a fraction of the cost, albeit slower.

The price difference is huge: the Intel card costs about $41 per gigabyte of video memory, while the NVIDIA card costs over $200 per gigabyte. For someone who wants to experiment with AI without breaking the bank, the Intel card is a compelling option. It's also good for other creative work like 3D rendering and simulations. However, getting the Intel card to work well requires technical know-how: you need to update the operating system kernel, enable a BIOS setting called Resizable BAR, and install specific software packages. Without these steps, the card's performance is disappointing.

The test was done on a custom-built computer with an older AMD Threadripper processor and 128GB of RAM, running Ubuntu Linux. The user notes that the NVIDIA performance numbers came from other people's tests, not their own machine, since they couldn't afford the expensive card. They also mention that if you can find an NVIDIA 5090 at its suggested retail price of $1,999, it would be a better deal, but current market prices are inflated.

For most people, this comparison shows that you don't need to spend a fortune to run AI locally. If you're willing to tinker and accept slower speeds, the Intel card can save you thousands of dollars. It's a great option for hobbyists and small businesses looking to explore AI without investing in top-tier hardware. Just be prepared for some setup challenges.