Gaudi 3: Intel launches next-generation GPUs for enterprise productive AI

Intel introduced its new Gaudi 3 generative artificial intelligence (AI) accelerators on Tuesday (9). four times higher computational performance than previous generation Gaudi 2. The device, introduced in Intel Vision 2024, aims to train advanced language models.

Line offers 1.5x more memory bandwidth and 2x more network bandwidthmaking it possible to power systems with tens of thousands of connected accelerators. It provides a significant leap in training capacity for companies looking to deploy generative AI at scale.

Manufactured at 5 nm, Intel Gaudi 3 includes 64 Tensor Processor Cores (TPC) and eight Matrix Multiplication Engines (MME) capable of performing 64 thousand parallel operations. The hardware also has 128 GB HBMe2 memory and 3.7 TB bandwidth. Processing large datasets with fewer GPUs.

Intel’s new generative AI accelerator also features Ethernet ports and open software for developer productivity. he proposes 50% faster training time than NVIDIA H100 Llama 7B, 70B and Falcon 180B models have 40% more energy efficiency.

More news on enterprise productive AI

Gaudi 3 cards will be available for manufacturers such as Dell, Hewlett Packard Enterprise (HPE), Lenovo and Supermicro in the second quarter of this year. In addition to next-generation GPUs, the company also announced others Solutions to accelerate the deployment of enterprise productive AI.

Among these Intel Xeon 6 processors The former with E or P cores offers 4x more performance and 2.7x better rack density, while the latter reduces next token latency by 6.5x compared to previous Xeon chips, allowing use with Llama 2 models with 70 billion parameters.

portfolio Intel TiberIt will help enterprise customers and developers find the best tools for their needs and the strategy for open and scalable AI systems. The approach will offer different solutions to improve services.

Companies that will benefit from the applications include Bosch, IBM, NAVER, IFF, Infosys, Landing AI, NielsenIQ and Seekr. Finally, collaborations with Google Cloud, Cohesity and Thales emerged. Intel Confidential Computing Resources in your cloud tools.