Intel-logo

Intel Gaudi 3

Ethernet-native AI accelerator designed for cost-efficient LLM training and high-throughput inference at scale.

Release

2024

GPU Class

Data Center / AI Accelerator

Architecture

Intel Gaudi (5th-gen heterogeneous architecture)

PRICE SNAPSHOT

Loading price comparison...
Loading live GPU prices...

On-premise Module

~$15k–$25k

Turnkey System

~$150k–$275k+

Cloud Pricing
(per GPU/hr)

~$5.50–$9.00/hr

chip identity

Intel-Gaudi-3

On-premise Module

Gaudi 3

GPU Class

Data Center / AI Accelerator

Release

2024

Architecture

Intel Gaudi (5th-gen heterogeneous architecture)

Target Workload

  • Large language model training and fine-tuning
  • High-throughput inference for text and multimodal generation
  • Scale-out distributed training and serving using Ethernet fabrics

Compatible Platforms

  • OEM 8-accelerator baseboards (OCP OAM 2.0) for data center deployments
  • PCIe add-in card deployments (enterprise and smaller-scale clusters)
  • Cloud AI instances (notably IBM Cloud offerings)
  • Interconnect: Integrated 24x 200GbE RDMA (RoCE v2) for scale-out
  • Software Stack: SynapseAI, Habana frameworks (PyTorch/DeepSpeed integrations), Hugging Face Optimum Habana

Ideal Buyer Profile

Enterprises and organizations requiring:

  • Cost-efficient scale-out training and inference using Ethernet fabrics
  • Flexibility to deploy on OEM systems without proprietary GPU interconnect dependencies
  • Willingness to adopt and operationalize the SynapseAI/Habana software ecosystem
  • A roadmap aligned with open networking, supply diversification, and large-cluster economics

Typical adopters include cloud providers offering alternative AI instances, enterprises building private AI clusters on Ethernet, and teams optimizing for throughput-per-dollar at multi-node scale.

Gaudi 3 is delivered through OEM platforms and select cloud providers rather than as a turnkey Intel-branded system. It is commonly deployed in 8-accelerator OAM baseboards for dense data center configurations and as PCIe add-in cards for broader server compatibility. Availability and pricing vary by OEM integration, cooling configuration, and volume, with procurement often tied to complete platform builds rather than single-card retail channels.

Availability Notes

Gaudi 3 is delivered through OEM platforms and select cloud providers rather than as a turnkey Intel-branded system. It is commonly deployed in 8-accelerator OAM baseboards for dense data center configurations and as PCIe add-in cards for broader server compatibility. Availability and pricing vary by OEM integration, cooling configuration, and volume, with procurement often tied to complete platform builds rather than single-card retail channels.

Recent Developments

  • Apr 2024 β€” Intel formally launches Gaudi 3 and details its Ethernet-native scale-out approach and HBM2e-based design.
  • Q3 2024 β€” Volume production and broader OEM platform availability ramp for dense OAM-based systems.
  • Early 2025 β€” IBM Cloud begins offering Gaudi 3 instances, positioning them as a lower-cost alternative for enterprise AI workloads.
  • 2025 β€” Continued software optimization work focuses on improving framework compatibility, inference performance, and operational stability in SynapseAI/Habana tooling.

overview

The Intel Gaudi 3 is a data center AI accelerator built for scalable training and inference with an architecture optimized around matrix engines, large on-package high-bandwidth memory, and integrated high-speed Ethernet networking. It targets generative AI workloads by combining strong BF16/FP8 performance with native scale-out connectivity, aiming to reduce reliance on proprietary GPU fabrics for multi-node scaling.

Gaudi 3 is delivered primarily as an OCP OAM 2.0 mezzanine module for dense 8-accelerator baseboards and as a PCIe add-in card for broader server compatibility. Its integrated 200GbE RDMA ports enable direct accelerator-to-accelerator networking, making it well suited to Ethernet-based AI clusters where cost, openness, and scale-out deployment patterns are primary considerations.

Key specifications

Specification Gaudi 3 AI Accelerator
Architecture Intel Gaudi (5th-gen heterogeneous architecture)
Memory 128Β GB HBM2e
Memory Bandwidth ~3.7Β TB/s
Interconnect 24xΒ 200GbE RDMA (RoCEΒ v2) integrated networking (platform-dependent usage)
Form Factor OCP OAMΒ 2.0 mezzanine (HL-325L), PCIe Gen5Β x16 add-in card (HL-338)
Max TGP Up to ~900Β W (OAM air-cooled); higher with liquid-cooled configurations; up to ~600Β W (PCIe)
Precision Support FP32, TF32-equivalent workflows (framework-dependent), FP16, BF16, FP8, INT8, FP64 (workload-dependent support)
Typical AI Compute ~1.8Β PFLOPS (FP8/BF16 peak, vendor-claimed)
Process Node TSMCΒ 5Β nm
Transistor Count TBD
MIG Support TBD (no NVIDIA MIG equivalent; partitioning/virtualization is platform and software dependent)
NVLink (peer) Not supported

Performance Summary

  • AI/ML Throughput: High FP8/BF16 throughput designed for LLM training and inference; performance advantages are most visible in scale-out configurations where Ethernet-based distribution is efficient and cost-sensitive.
  • Tensor Compute: Dedicated matrix engines and software stack optimizations target transformer-heavy workloads, with vendor-claimed peak compute around ~1.8 PFLOPS (FP8/BF16).
  • Memory Bandwidth: 128 GB HBM2e and ~3.7 TB/s bandwidth help reduce memory bottlenecks for larger batch sizes and higher-throughput inference serving.
  • Multi-Accelerator Scaling: Integrated 24x 200GbE RDMA enables native Ethernet scale-out without a proprietary GPU fabric, shifting scaling efficiency toward network design, topology, and distributed software tuning.

Compared to NVIDIA H100-class systems, Gaudi 3’s positioning emphasizes scale-out economics and open networking alongside competitive LLM throughput on supported frameworks, while ecosystem maturity and software portability remain key decision factors.

primary use case

  • Large-scale LLM training and fine-tuning on Ethernet-based clusters
  • High-throughput inference for generative AI services where cost-per-token is a primary metric
  • Enterprise AI deployments prioritizing open scale-out networking and supply-chain diversity
  • Distributed training workloads that scale effectively across multiple nodes using RDMA networking
  • Multimodal inference and batch serving pipelines supported by optimized framework integrations

Alternatives & Upgrade Path

Comparable AI accelerators:

  • NVIDIA H100: Hopper-generation baseline with strong scale-up via NVLink and broad CUDA ecosystem maturity.
  • NVIDIA H200: Hopper refresh optimized for higher memory capacity and bandwidth; often preferred for memory-bound LLM inference and long-context workloads.
  • AMD Instinct MI300X: High-memory accelerator (192 GB HBM3) positioned for large-model inference and memory-heavy workloads.

For organizations prioritizing open networking and scale-out economics, Gaudi 3 is typically evaluated against H100/H200-class deployments on a cost-per-throughput basis. Buyers requiring maximal software portability or tightly coupled multi-GPU scale-up may favor NVIDIA platforms, while those optimizing for Ethernet-native clusters may view Gaudi 3 as a scalable alternative.

Related Chips & Providers

Intel accelerators:

  • Gaudi 2 (prior generation)
  • Intel AI accelerators supported via SynapseAI software stack

Competitor GPUs:

  • NVIDIA H100
  • NVIDIA H200
  • AMD Instinct MI300X

SUMMARY

Intel Gaudi 3 is a data center AI accelerator optimized for cost-efficient LLM training and inference with an Ethernet-native scale-out model. It combines 128 GB of HBM2e, ~3.7 TB/s memory bandwidth, and integrated 24x 200GbE RDMA networking to support large distributed deployments without a proprietary GPU fabric. Gaudi 3 is typically deployed in dense 8-accelerator OEM platforms or PCIe form factors, appealing most to organizations prioritizing open networking, throughput-per-dollar, and an alternative software stack to CUDA.

Newsletter

Stay Ahead in Cloud
& Data Infrastructure

Get early access to new tools, insights, and research shaping the next wave of cloud and storage innovation.