Intel Unveils the Most Efficient Gaudi 3 AI Accelerator at Intel Vision

Intel has announced its latest AI chip Gaudi 3 at the Intel Vision 2024 event, in a bid to keep pace with the growing demand for semiconductors capable of training and deploying large AI models.

The newly introduced Gaudi 3 chip, which was revealed by CEO Pat Gelsinger at Intel AI Everywhere event, boasts over double the power efficiency compared to its predecessor and is capable of running AI models 1.5 times faster than NVIDIA’s H100 GPU. 

It offers various configurations, including a bundle of eight Gaudi 3 chips on one motherboard or a card that can be integrated into existing systems.

Gaudi 3, built on a 5 nm process, signals Intel's utilisation of advanced manufacturing techniques. Additionally, Intel plans to manufacture AI chips, potentially for external companies, at a new Ohio factory expected to open in the coming years, according to Gelsinger.

During testing, Intel evaluated the chip's performance on models like Meta’s open-source Llama and the Falcon model by TII. Gaudi 3 demonstrated its capability to facilitate the training or deployment of models such as Stable Diffusion or OpenAI’s Whisper model for speech recognition.

Compared to the NVIDIA H100, Intel anticipates that Gaudi 3 will achieve an approximately 50% faster time-to-train on average across Llama 2 models with 7B and 13B parameters, as well as the GPT-3 175B parameter model. 

While performance data for NVIDIA’s recently announced Blackwell-based B200 Tensor GPU is not currently available, it's clear that NVIDIA's latest offering would likely affect these performance comparisons significantly.

In comparison to NVIDIA, Intel claims its chips consume less power. NVIDIA currently dominates approximately 80% of the AI chip market with its GPUs, which have been the preferred choice for AI developers in the past year.

Intel asserts that its Gaudi 3 AI accelerator offers an estimated 50% enhancement in inferencing performance and around 40% better power efficiency compared to NVIDIA's H100. Moreover, Intel states that it achieves these benefits at a significantly lower cost.

Intel has announced that Gaudi 3 chips will be available to customers in the third quarter, with companies like Dell, HPE, and Supermicro set to incorporate the chips into their systems. However, Intel hasn’t disclosed the pricing details for Gaudi 3.

Das Kamhout, vice president of Xeon software at Intel, expressed confidence in Gaudi 3’s competitiveness against NVIDIA’s latest offerings, citing factors such as competitive pricing and the incorporation of an open integrated network on chip.

The data centre AI market is expected to expand as cloud providers and businesses invest in infrastructure for deploying AI software, indicating opportunities for other players in the market.

While NVIDIA has seen significant stock growth driven by the AI boom, Intel's stock has experienced more modest gains. Nevertheless, Intel remains determined to compete in the AI chip market, with AMD also seeking to expand its presence in the server AI chip segment.

NVIDIA’s success has largely been attributed to its proprietary software suite, CUDA. In contrast, Intel is collaborating with chip and software giants like Google, Qualcomm, and Arm to develop open software solutions, aiming to provide greater flexibility for software companies in selecting chip providers.

In addition, Intel unveiled its intention to create an open platform for enterprise AI, aiming to expedite the deployment of secure GenAI systems empowered by retrieval augmented generation (RAG).