Intel Unveils Cost-Effective Gaudi AI Chips to Compete with Nvidia

Intel has announced its new line of Gaudi AI chips, designed to offer high-performance artificial intelligence processing at a reduced expense compared to rival Nvidia's offerings. The move comes as Intel aims to capitalize on the growing demand for AI technology across various industries and challenge Nvidia's dominance in the AI chip market.

The Gaudi 3 AI chip, Intel claims, is 40% more power-efficient than Nvidia's H100 GPU. This development occurs as tech giants like Google, Intel, and Meta are all developing their own AI chips to reduce reliance on Nvidia's expensive AI solutions. Google has introduced its Axiom Processors, while Meta has announced its new MTIA (Meta Training and Inference Accelerator) chip, offering three times the performance of its previous-generation chip.

Intel has also collaborated with Meta to ensure its AI hardware and software products are optimized for the latest large language models, including Meta's Llama 3 models. The Gaudi 2 accelerators have been optimized for Llama 2 models and now have initial performance measurements for the new Llama 3 model. Intel Xeon processors and Intel Core Ultra processors also show impressive performance for running the Llama 3 models.

Why this matters: The rise of AI has posed challenges for journalists, including ethical and editorial concerns, as AI tools are being used to assist or replace some newsroom tasks. Media organizations are developing their own AI tools and grappling with issues like content provenance and authenticity, while regulation of AI in the media industry remains in its early stages.

The Gaudi 3 chip features 128GB of HBM2e memory, up to 1.835 PFLOPS of FP8 compute, and 64 tensor processor cores. Built on a 5nm process, it is a notable improvement from the 7nm Gaudi 2. The Gaudi 3 uses Ethernet as the interconnect between accelerators, allowing for easier scaling in data centers compared to NVIDIA's proprietary NVLink and NVSwitch.

While Intel claims the Gaudi 3 is more power-efficient and sometimes faster than NVIDIA's H100 in inferencing, NVIDIA's upcoming H200 and Blackwell products may offer faster training performance. Intel is targeting the AI inference market with the Gaudi 3 and also offering a 600W PCIe add-in card version, the Gaudi HL 338. The Gaudi 3 accelerators are sampling in the first half of 2024.

Intel needs to prove that its R&D, production, delivery, and partner initiatives are back on track after several years of missteps, as its revenue declined 14% in 2023 compared to Nvidia's 126% revenue growth. Intel is expected to announce its Q1 2024 earnings on April 25, while Nvidia's Q1 2024 earnings will likely arrive in May.

The biggest potential winner from these AI chip launches may be Supermicro Computer, a leader in the fiel

Key Takeaways

• Intel announces new Gaudi AI chips to challenge Nvidia's dominance.
• Gaudi 3 chip claims 40% more power-efficiency than Nvidia's H100 GPU.
• Intel collaborates with Meta to optimize Gaudi chips for Llama 3 models.
• Gaudi 3 features 128GB HBM2e, 1.835 PFLOPS FP8 compute, and 64 tensor cores.
• Intel aims to regain market share after revenue decline, while Nvidia sees growth.