Intel has made headlines by integrating its Gaudi 3 rack-scale solution with NVIDIA’s advanced technology stack. This innovative combination leverages Intel’s own AI chips alongside NVIDIA’s Blackwell GPUs, promising substantial performance enhancements for users.
Intel Unveils a Hybrid AI Server with NVIDIA’s Blackwell Technology
The Gaudi line of AI chips from Intel has gained significant traction in the industry. However, the company has faced stiff competition from giants like NVIDIA and AMD in capturing revenue from the burgeoning AI sector. To address this challenge, Intel is reimagining its strategy for the Gaudi platform. As reported by SemiAnalysis, the tech giant is set to introduce the Gaudi 3 rack-scale system, which incorporates NVIDIA’s Blackwell B200 GPU as part of a hybrid architecture, complemented by Connect-X networking technology.
Intel just took another step on combining forces 🔥 with NVIDIA by integrating their new Gaudi3 rack scale systems together with NVIDIA B200 via disaggregated PD inferencing. Intel claims that compared their B200 only baseline, and inferencing system using Gaudi3 for decode part… pic.twitter.com/jAKin6rgZx
— SemiAnalysis (@SemiAnalysis_) October 18, 2025
This announcement stands out as one of the key highlights from the recent OCP Global Summit, where Intel aims to carve out a distinctive niche within the rack-scale AI market. The proposed system uniquely employs Intel’s Gaudi 3 chips to handle the ‘decode’ aspects of inferencing workloads, while the B200 GPUs focus on the more demanding ‘prefill’ phases. Blackwell GPUs are characterized by their exceptional performance in large matrix-multiplications, making them an optimal choice for handling prefill operations.
In this innovative configuration, the Gaudi 3 architecture will prioritize memory bandwidth and Ethernet-centric scalability. On the connectivity front, the setup utilizes NVIDIA’s ConnectX-7 400 GbE NICs mounted on compute trays, alongside Broadcom’s Tomahawk 5 switches, boasting a remarkable 51.2 Tb/s throughput to facilitate full rack connectivity. According to SemiAnalysis, each compute tray is equipped with two Xeon CPUs, four Gaudi 3 AI chips, and four NICs, in addition to a NVIDIA BlueField-3 DPU, with a total of sixteen trays allocated per rack.
The Gaudi platform is positioned as a cost-effective decoding engine amidst a landscape heavily influenced by NVIDIA’s dominance. This strategy suggests a pragmatic approach where Intel seeks to enhance its market standing not by direct competition, but by leveraging a collaborative arrangement. Claims suggest that this rack-scale architecture can achieve 1.7 times faster performance in prefill tasks compared to a baseline model using only the B200 GPU; however, these results await independent validation.
While this hybrid setup presents an optimistic future, challenges remain. The Gaudi platform is still hindered by an undeveloped software ecosystem, which may impede its broader adoption. Furthermore, with the Gaudi architecture scheduled for phased retirement in the coming months, it’s uncertain whether this configuration will achieve the same mainstream acceptance as competing solutions.
Related Articles:
“Hey Copilot: A Nostalgic Look at Cortana”
16:32
ASRock RX 9070 XT Monster Hunter Wilds Edition Now Available on Newegg for $699
14:28
NVIDIA Champions ‘Made in USA’ as Jensen Huang Reveals First Blackwell Chip Wafer Manufactured by TSMC in Arizona
8:22
Leave a Reply