Intel Gaudi AI Accelerator
First-generation deep learning training and inference for competitive price/performance
Click here to jump to more pricing!
For leading price performance in the cloud and on-premises
With the first-generation Intel Gaudi AI deep learning processor, customers benefit from the most cost-effective, high-performance training and inference alternative to comparable GPUs. This is the deep learning architecture that enables AWS DL1 instances based on the Intel Gaudi AI accelerator to deliver up to 40% better price/performance training as compared to comparable Nvidia GPU-based instances. Intel Gaudi AI accelerator’s efficient architecture also enables Supermicro to provide customers with equally significant price performance advantage over GPU-based servers with the Supermicro X12 Server featuring Intel Gaudi AI accelerators.
What makes the Intel Gaudi AI accelerator so efficient?
16nm
Process technology
DL Optimized
Matrix multiplication engine
8 Programmable
Tensor processor cores
32 GB
Onboard HBM2
24 SRAM
10 Integrated
100G Ethernet ports
Massive and flexible system scaling with Intel Gaudi AI accelerator
Every first-generation Intel Gaudi AI processor integrates ten 100 Gigabit Ethernet ports of RDMA over Converged Ethernet (RoCE2) on chip to deliver unmatched scalability, enabling customers to efficiently scale AI training from one processor to 1000s to nimbly address expansive compute requirements of today’s deep learning workloads.
Options for building Intel Gaudi AI accelerator systems on premises
For customers who want to build out on-premises systems, we recommend the Supermicro X12 Server, which features eight Intel Gaudi AI processors. For customers who wish to configure their own servers based on Intel Gaudi AI accelerators, we provide reference model options, the HLS-1 and HLS-1H.
Making developing on Intel Gaudi AI accelerators fast and easy: Intel Gaudi Software
Optimized for deep learning model development and to ease migration of existing GPU-based models to Intel Gaudi AI platform hardware. It integrates PyTorch and TensorFlow frameworks and supports a rapidly growing array of computer vision, natural language processing and multi-modal models. In fact, over 200K models on Hugging Face are easily enabled on Intel Gaudi accelerators with the Habana Optimum software library. Getting started with model migration is as easy as adding 2 lines of code, and for expert users who wish to program their own kernels, the Intel Gaudi platform offers the full tool-kit and libraries to do that as well. Intel Gaudi software supports training and inference of models on first-gen Intel Gaudi accelerators and Intel Gaudi 2 accelerators.
