Intel launches new GPUs focused on AI inference and workstations, and also brings updates for its Gaudi 3 AI accelerators.
At Computex, Intel expands its portfolio of GPUs “and accelerators”. The manufacturer introduces the Arc Pro B60 and B50 GPUs. The chips are aimed at professional workstations and, according to Intel, built with AI workloads in mind.
Diverse Target Group
Both models are based on the Xe2 architecture and offer features such as Intel Xe Matrix Extensions (XMX) AI cores and advanced ray tracing. They are intended for applications in architecture, engineering, and construction, as well as AI inference tasks. The GPUs are supported by various independent software vendors and have compatible drivers for Windows and Linux.
The B60 has 24 GB of memory while the B50 makes do with 16 GB. In terms of AI, that’s not very impressive. Certainly, 16 GB is not enough to load large LLMs. The Arc Pro B50 thus seems more like a general workstation GPU.
Additionally, Intel unveils a workstation platform based on Intel Xeon, codenamed Project Battlematrix. This platform can contain up to eight Arc Pro B60 GPUs with a total video memory capacity of 192 GB. This allows running medium-sized AI models with up to 150 billion parameters.
More Intel Gaudi 3
Intel also expands its AI accelerator offerings with more Gaudi 3 options. The accelerator is now available as a PCIe card and in rack scale systems. The PCIe cards support scalable AI inference and are suitable for both small businesses and large enterprises. They can handle various AI models, including Llama 3.1 8B to Llama 4 Scout and Maverick. The cards will be available in the second half of 2025.
The rack scale solutions can contain up to 64 accelerators per rack and have 8.2 TB of high-bandwidth memory. Intel opts for an open and modular design independent of a single vendor.
Tough Competition
Intel is trying to grab a piece of that lucrative market with its GPUs and accelerators. This is challenging, as the entire AI sector is dominated by Nvidia. That manufacturer not only has the most capable chips but also supports them with a software ecosystem that is essentially the standard today.
Intel still has a high market share in servers and workstations, especially with Xeon. Combined solutions are certainly relevant, although the manufacturer is still lacking some momentum here. The underdog position has already been taken by AMD, which is trying to offer an interesting alternative to Nvidia with its Epyc chips and Instinct accelerators. In that respect, AMD has built up more credibility, partly due to the use of its accelerators in the world’s most powerful supercomputers.