Shulou(Shulou.com)11/24 Report--

Today, MLCommons released the results of its industry AI performance benchmark MLPerf training 3.0, in which the Habana ®Gaudi ®2 Deep Learning Accelerator and the fourth generation Intel ®strong ®scalable processors demonstrated impressive training results.

Sandra Rivera, Executive Vice President and General Manager of data Center and artificial Intelligence at Intel, said: "the latest MLPerf results released by MLCommons demonstrate that using Intel Xeon scalable processors and Intel Gaudi Deep Learning Accelerator can bring customers higher performance-to-price ratio (TCO) in AI. Among them, the powerful built-in accelerator makes it an ideal solution for running a large number of AI workloads on general-purpose processors, while Gaudi provides competitive performance for large language models and generative AI. In addition, Intel's scalable system is equipped with optimized, easy-to-program open software that lowers the bar for customers and ecological partners to deploy AI-based solutions from the cloud to the smart edge in the data center. "

At present, it is generally believed that generative AI and large language model (LLMs) are only suitable for running on GPU. However, the latest data show that AI solutions based on Intel's portfolio can provide competitive options for customers seeking to get rid of current efficiency and scale constraints in a closed ecosystem.

The latest MLPerf training 3.0 results demonstrate the excellent performance of Intel products on a range of deep learning models. On the large language model GPT-3, the software and system based on Gaudi2 have been verified on a large scale in the maturity of AI training. It is worth mentioning that Gaudi2 is one of only two solutions that submit performance results to the GPT-3 large model training benchmark.

At the same time, Gaudi2 also provides customers with competitive cost advantages, including server and system costs. Its outstanding performance proven by MLPerf on GPT-3, computer vision and natural language models, as well as upcoming software, make Gaudi2 an attractive and cost-effective solution in the industry.

In terms of CPU, the fourth generation Xeon scalable processor uses Intel AI engine. The results of its deep learning training performance show that customers can use Xeon servers to build a general AI system for data preprocessing, model training and deployment, so as to obtain the optimal combination of AI performance, efficiency, accuracy and scalability.

Test results for Habana Gaudi2: training generative AI and large language models require server clusters to meet large-scale computing requirements. The latest MLPerf results really verify the excellent performance and efficient scalability of Habana Gaudi2 on the extremely demanding model-GPT-3 with 175 billion parameters.

Test highlights:

Gaudi2 achieved an impressive training time on GPT-3 *: 311 minutes on 384 accelerators.

On the GPT-3 model, 95% near-linear expansion is achieved from 256accelerators to 384accelerators.

Excellent training results have been obtained on the computer vision model ResNet-50 (8 accelerators) and Unet3D (8 accelerators) and the natural language processing model BERT (8 and 64 accelerators).

Compared with the data submitted in November last year, the performance of the BERT and ResNet models improved by 10 per cent and 4 per cent respectively, proving an increase in the maturity of Gaudi2 software.

Gaudi2 supports the "out of the box" feature, and customers can get similar performance results to this test when using Gaudi2 locally or in the cloud.

Habana ®Gaudi ®2 mezzanine card

On the software maturity of Gaudi2: Gaudi's software support continues to grow and mature, and can keep pace with the growing demand for generative AI and large language models.

The GPT-3 model submitted this time is based on PyTorch and uses the current popular DeepSpeed optimization library belonging to Microsoft's large-scale AI, rather than customized software. DeepSpeed can support 3D parallelism of Data, Tensor and Pipeline at the same time, which further optimizes the extended performance efficiency of large language models.

The Gaudi2 results of MLPerf 3.0 have been submitted as BF16 data type. Gaudi2 performance is expected to take a significant leap when software support and new features for FP8 are released in the third quarter of 2023.

Test results for fourth-generation Xeon scalable processors: as the only CPU-based solution submitted among many solutions, MLPerf results show that Intel Xeon scalable processors provide enterprises with "out-of-the-box" functionality to deploy AI on general-purpose systems, avoiding the high cost and complexity of introducing dedicated AI systems.

For a small number of users who train large models intermittently from scratch, they can use generic CPU and usually run their business on Intel-based servers that have been deployed. In addition, most people will use pre-trained models and fine-tune them with small data sets. The results released by Intel show that this fine-tuning can be done in just a few minutes by using Intel AI software and standard industry open source software.

Highlights of MLPerf testing:

In the closed area, the fourth generation of the strongest can train the BERT and ResNet-50 models in less than 50 minutes (47.93 minutes) and 90 minutes (88.17 minutes) respectively.

For the open area of the BERT model, the results show that the fourth generation Xeon can complete the model training in about 30 minutes when it is extended to 16 nodes.

For larger RetinaNet models, the fourth-generation Xeon can achieve 232 minutes of training time on 16 nodes, giving customers the flexibility to use off-peak Xeon cycles to train their models, that is, in the morning, lunch or night.

The fourth-generation Intel Xeon scalable processors with Intel ®Intel ®AMX provide significant "out-of-the-box" performance improvements, covering multiple frameworks, end-to-end data science tools, and a broad ecosystem of smart solutions.

Fourth-generation Intel ®strong ®Extensible processor

MLPerf is widely considered to be the most convincing benchmark for AI performance testing, allowing fair and repeatable performance comparisons between various solutions. At present, Intel has more than 100 performance results and is the only vendor that uses industry-standard deep learning ecosystem software and publicly submits CPU results.

The results also show that excellent scalability can be achieved using cost-effective and readily available Intel Ethernet 800 series network adapters using an open source Intel ®Ethernet software package based on Intel oneAPI.

Description:

* the MLPerf test corpus consists of 1% of the GPT-3 model representatives.

Declaration:

Product performance may vary depending on how it is used, configuration, and other factors. For more information, visit www.Intel.com/ PerformanceIndex.

Performance results are based on tests as of the date shown in the configuration and may not reflect all publicly available updates. No product or component is absolutely safe.

Your cost and performance results may vary.

Intel technology may need to be activated by enabling hardware, software, or services.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.