Specifications

Overview

The Intel Gaudi 3 AI Accelerator (HL-338) is Intel's third-generation deep learning processor, delivering 1,835 TFLOPS of FP8 AI compute in a standard PCIe Gen 5.0 add-in card form factor. Built on a 5nm process with 128 GB of HBM2e memory at 3.7 TB/s bandwidth, Gaudi 3 is purpose-built for large-scale AI training and inference with an open, Ethernet-native architecture.

Unlike proprietary GPU interconnects, Gaudi 3 uses standard 200GbE RoCE v2 RDMA with 24 integrated ports — eliminating the need for InfiniBand or NVLink. Each accelerator connects directly to Ethernet switches, enabling cost-effective scale-out to thousands of nodes. For scale-up, 4-card quads deliver 900 GB/s of inter-card bandwidth with 512 GB of pooled HBM2e memory, and 8-card configurations double that to 1,024 GB.

Gaudi 3's 64 TPC (Tensor Processor Cores) and 8 MME (Matrix Multiplication Engines) are natively designed for deep learning workloads including large language models, diffusion models, and multi-modal AI. The dedicated media processor offloads image/video decode and pre-processing. Available in PCIe, mezzanine (HL-325L), and UBB (HLB-325) form factors, Gaudi 3 integrates with Dell PowerEdge XE7740 and other OEM server platforms. QS Compute can source Intel Gaudi 3 accelerators for enterprise AI infrastructure deployments.