Intel today's second-generation Habana® Gaudi®2 deep learning processor and NVIDIA A100's AI total learning time (TIME-TO-Train, TTT) performance on MLPERF Industrial Benchmark The processor's performance was superior. Intel said in May that the Gaudi 2 processor, which was announced in Intel Vision, recorded excellent TTT in the Vision (Resnet-50) and Language (BERT) sectors.

Sandra Rivera, Intel's senior vice president and data center, and AI Group, said, Gaudi 2 has excellent performance in the MLPERF benchmark within a month of its launch, and at the same time, we are proud of our team members who have tried to achieve these results at the same time. I feel it, he said. Intel will provide the best performance in both vision and language models to provide customers with value and accelerate the development of AI deep learning solutions.

The Intel Data Center team focused on deep learning processor technology by utilizing Gaudi platforms of Habana Labs, and supported data scientists and machine learning engineers to accelerate learning. In addition, the company has established a new model with just a few lines of code, or relocates existing models to increase productivity and reduce operating costs.

Gaudi 2, Havana, has made a breakthrough in the TTT sector compared to the first generation Gaudi. Havana Labs said at the MLPERF benchmark in May 2022 that Gaudi 2 was excellent for the NVIDIA A100-80G in the vision and language model that uses eight accelerators. For the Resnet-50 model, Gaudi 2 has a 36% reduction in learning time compared to the NVIDIA A100-80G product. Dell (DELL) 's tests of the Resnet-50 and BERT model learning tests on eight accelerators servers showed shortened the learning time compared to the NVIDIA A100-40GB.

Gaudi 2 recorded three times and 4.7 times higher than the first-generation Gaudi, Resnet-50 and BERT models, respectively. Intel converted the processor from the existing 16-nano process to the 7nm process, increasing the number of tensor processors cores, and expanding the GEMM engine computing capacity, expanding triples in the package in the package, and doubling the size of bandwidth and SRAM. This achievement was achieved. In the case of the Vision Model, Gaudi 2 is independently operated and has an integrated media engine type function that can handle overall pretreatment pipes for compressed imaging, including data enhancement required for AI learning.

Gaudi 1 and Gaudi 2 processors provide the best performance to customers without special software manipulation.

Havana Labs compares the performance between Gaudi 1 and Gaudi 2 and existing commercial software on eight GPU servers and HLS-Gaudi 2 reference servers. The learning process was measured using the Tensorflow Docker of the NGC and the Havana Public Repository, and adopted the best performance parameters recommended by the manufacturer.

In addition to the performance of Gaudi 2 products measured through MLPERF, Gaudi 1 provided powerful performance and linear scale in the 128-accelerator that supports high-efficiency system scaling and the Resnet model for 256-accelerator.

Gaudi 2 provides industry leading performance in this model learning, as Gaudi 2 has proved as a result of the latest MLPERF results, said Eitan Medina Habana Labs. We are continuously innovating deep learning education architecture and software to provide.

As for MLPERF benchmarks: The MLPERF community aims to design a fair and useful benchmark that provides consistent measurement of accuracy, speed and efficiency for machine learning solutions. Leaders in academia, lab, and industry have defined a series of strict rules that determine the benchmark and guarantee fair comparison between all suppliers. The MLPERF benchmark is the only benchmark that can be trusted in the AI industry due to the explicit set set that enables fair comparison of the end-to-end work. In addition, if you submit the results to the MLPERF, you will go through a colleague review process for a month to verify it.