Intel® Gaudi® 3 Platform with GIGABYTE solutions

Leap forward in performance and efficiency with an open, Ethernet-based AI system.

Performance and Efficiency at Every Scale

Building on extensive experience in accelerator design and deep expertise in microarchitecture and software, Intel has developed the third generation of Gaudi AI accelerators, Intel^® Gaudi^® 3, delivering breakthrough performance and efficiency. These accelerators achieve competitive results that rival leading solutions while maintaining a scalable platform. With a strong emphasis on Ethernet adoption and an open system architecture, Intel Gaudi 3 AI accelerators set a new standard for AI infrastructure, enabling businesses to scale efficiently and meet the evolving demands of tomorrow's AI challenges.

Scaling AI with GIGABYTE Server on Intel Gaudi 3 Solution

Always striving for the perfect balance between performance, efficiency, stability, and scalability, GIGABYTE has developed numerous designs to fit various use cases for these AI-era GPU powerhouses. For this newcomer to the GIGABYTE AI lineup, a robust 8U chassis with optimized thermal capabilities was designed to extract every ounce of its potential. It marks the first GIGABYTE server to adopt an 8U air-cooling solution that seamlessly fits into industry-standard air-cooled infrastructure.

By fully utilizing this Ethernet-centric, scalable solution on GIGAPOD – GIGABYTE’s well-optimized and proven rack solution – customers can quickly adopt the latest Intel Gaudi 3 solution with minimal verification required. The rack solution features a 4-server configuration with a Rear Door Heat Exchanger (RDHx), maximizing compute density for optimal utilization of limited facility space.

To learn more about GIGAPOD, please visit our GIGAPOD solution page.

Designed for the Real-World Demands of AI

Adopt with Ease

Effortless adoption or migration of existing code with Intel Gaudi software, purpose-built for Gen AI with industry-leading software capabilities.
Built with Scalability

Designed for Ethernet hardware with 1200GB/s open standard RoCE connection among accelerators, scaling cost-effectively for even the largest and most complex deployments.
Flexible and Powerful Computing

A mix of 8 Matrix Multiplication Engines (MME) and 64 Tensor Processor Cores (TPC) on two interconnected compute dies, delivering optimal performance across a wide range of workloads.
Efficient Memory Intensive Computing

A total of 128GB of HBM and 96MB L2 cache, effectively addressing the memory bottlenecks often seen in AI training and inference, efficiently accelerating memory-intensive applications like LLM.

Intel Gaudi 3 AI Accelerator Specifications

Model	Intel^® Gaudi^® 3 Accelerator
BF16/FP8 MME TFOPs	1835
BF16 Vector TFLOPs	28.7
MME Units	8
TPC Units	64
HBM Capacity	128 GB
HBM Bandwidth	3.7 TB/s
On-die SRAM Capacity	96 MB
On-die SRAM Bandwidth	12.8 TB/s
Networking	1200 GB/s bidirectional
Host Interface	PCIe Gen5 x16
Media	14 Decoders