By Timothy Green
Publication Date: 2026-03-27 10:30:00
Intel (INTC 6.53%) just announced a new GPU. It’s not for gaming, and it’s not destined for data centers. Instead, the Intel Arc Pro B70 is tailor-made for running AI workloads on workstations locally.
Priced at $949, the Arc Pro B70 is drastically less expensive than the $1,800 RTX Pro 4000 from Nvidia and the $1,299 Radeon AI Pro R9700 from AMD. While Nvidia retains an important advantage with CUDA, its proprietary software platform that protects its massive market share, Intel’s claimed performance advantage could be enough for the company to make some inroads in this market.
Image source: Intel.
Beating Nvidia at its own game
Frontier AI models require enormous quantities of memory and computing power, so they can only be run in massive data centers with high-powered AI accelerators. However, not every AI workload requires Claude Opus or GPT-5.4. There are plenty of use cases that can be satisfied with smaller AI models that fit on a workstation GPU.
The Arc Pro B70 has 32 Xe cores that deliver 22.9 TFLOPS of FP32 compute performance. It features 32GB of GDDR6 RAM, enabling 608 GB/s bandwidth. This type of memory is less performant than the high-bandwidth memory typically used in data center GPUs, but it’s also cheaper. Intel claims the Arc Pro B70 can achieve 367 TOPS (trillions of operations per second).
The high memory content enables running a wider range of AI models with larger contexts. Using the open-source Llama 3.1 8B model, Intel claims that the Arc Pro…
