Intel's Habana® Gaudi® 2 deep learning processor has surpassed NVIDIA A100 in MLPerf benchmarks, demonstrating best-in-class performance.

Intel announced on the 1st that when measuring the AI total training time (Time-to-Train, hereafter TTT) performance of its second-generation Habana® Gaudi® 2 deep learning processor against NVIDIA A100 on the MLPerf industry benchmark, the performance of Habana® Gaudi® 2 deep learning processor proved superior.

Intel stated that the Gaudi2 processor announced at Intel Vision last May achieved outstanding TTT in vision (ResNet-50) and language (BERT) categories.

Sandra Rivera, Intel Senior Vice President and General Manager of Data Center and AI Group, said "I am delighted that Gaudi2 achieved outstanding performance in the MLPerf benchmark just one month after launch, and I am proud of the team members who made this result possible. Intel will provide best-in-class performance in both vision and language models, delivering value to customers and accelerating the development of AI deep learning solutions."

Intel's Data Center team focused on deep learning processor technology utilizing Habana Labs' Gaudi platform and supported data scientists and machine learning engineers to accelerate training. Additionally, it enabled building new models or migrating existing models with just a few lines of code, increasing productivity and reducing operational costs.

Habana Gaudi2 achieved breakthrough progress in the TTT category compared to the first-generation Gaudi product.

Habana Labs announced that through MLPerf benchmarks conducted in May 2022, Gaudi2 recorded superior performance compared to NVIDIA A100-80G in vision and language models using eight accelerators.

For the ResNet-50 model, Gaudi2 shortened training time by 36% compared to the NVIDIA A100-80G product.

In training tests of ResNet-50 and BERT models conducted on eight-accelerator servers by Dell, Gaudi2 reduced training time by 45% compared to NVIDIA A100-40GB.

Gaudi2 recorded 3 times and 4.7 times higher training throughput compared to first-generation Gaudi in ResNet-50 and BERT models, respectively.

Intel transitioned the processor from the existing 16-nanometer process to a 7-nanometer process, tripling the number of tensor processor cores, and achieved this result through expanded GEMM engine computing capacity, a threefold increase in high-bandwidth memory capacity within the package, and doubled bandwidth and SRAM size.

In the case of vision models, Gaudi2 operates independently and features integrated media engine functionality that can handle the entire preprocessing pipeline for compressed image visualization, including data augmentation required for AI training.

Gaudi1 and Gaudi2 processors deliver best performance to customers without requiring special software manipulation.

Habana Labs compared the performance between Gaudi1, Gaudi2, and conventional commercial software on eight-GPU servers and HLS-Gaudi2 reference servers. Training throughput was measured using TensorFlow Docker from NGC and Habana public repositories, and optimal performance parameters recommended by manufacturers were adopted.




Beyond the Gaudi2 performance measured through MLPerf, Gaudi1 provided strong performance and linear scaling in ResNet models for 128-accelerator and 256-accelerator systems supporting high-efficiency system scaling.

Eitan Medina, Chief Operating Officer of Habana Labs, stated "Gaudi2 delivers industry-leading performance in model training as proven by the latest MLPerf results. Habana Labs continues to innovate deep learning training architecture and software to provide cost-competitive AI training solutions."