Intel reveals its AI strategic blueprint: Gaudi 3 will be released this year, and Falcon Shores and other products will be launched next year

Accelerate AI layout

In terms of artificial intelligence, Intel has developed a complete enterprise AI stack (Enterprise AI stack), using open source standards such as OpenVINO and Intel Developer Cloud oneAPI & Synapse to promote the rapid development of hardware, systems, and applications.

Hardware is divided into three branches: Data Center (scalable systems, accelerators, CPUs), Networking (open standards and configurability, infrastructure), and Client & Edge (AI PCs, NPUs, GPUs, CPUs).

Gaudi 2 is more cost-effective

Intel said that in terms of pricing, Gaudi 2 uses the same process as Nvidia A100, which is TSMC's 7nm process, but the key difference is that in terms of Stability Diffusion and generative AI loads, Gaudi 2's performance is 3 times that of A100, which means Using Gaudi 2 can save more costs than Nvidia's A100. Gaudi 2 hardware also outperforms the latest H100 GPUs in certain workloads, and Intel says it will continue to expand on these advantages to maintain its lead in the future.

Gaudi 3 is scheduled to be released in the third quarter of 2024

In terms of roadmap, Intel is still positioning Gaudi 3 as the follow-up to Gaudi 2 and Falcon Shores. Intel said the Gaudi 3 accelerator is currently in validation labs and will be generally available in the third quarter of 2024.

Intel said that in terms of software/applications, it has built an extensive OEM ecosystem and Intel Developer Cloud Network for Gaudi 3. In terms of hardware, the Gaudi 3 accelerator will use the same high-performance architecture as Gaudi 2, but has 4 times the computing power. , the network bandwidth is 2 times, and the HBM memory bandwidth is 1.5 times.

Falcon Shores plans to launch in 2025

Falcon Shores is the first GPU architecture designed specifically for artificial intelligence workloads. It combines the best of Gaudi AI IP and Intel GPU technology into a converged product. Development of Falcon Shores is progressing well, with a launch targeted for 2025.

Model migration can be achieved in 3-5 lines of code

A major concern among AI developers is that migrating to new hardware will pose problems, such as modifying code to support new accelerators. According to Intel, most developers working at the higher levels of the stack (Hugging Face, Mosiac ML, PyTorch, etc.) can leverage Intel's software solutions to migrate existing applications by simply adding a few lines of code to a Python script. There are models or new models.Intel's "Model Migration" requires only 3-5 lines of code to migrate existing or new models to new hardware. For advanced software, Intel offers migration support from NVIDIA to the entire Gaudi and even next-generation Falcon Coast accelerator lineup.The company is also working to add low-level models such as OpenAPI, Triton and TPC-C to provide support when migrating from NVIDIA hardware to Falcon Shores GPUs.

China-available version of Gaudi hardware to be launched

Intel said it will customize Gaudi hardware for the Chinese market so that it can be exported to the Chinese market in full compliance with U.S. authorities.