While the artificial intelligence (AI) accelerator market is dominated by Nvidia (NVDA -0.47%), Intel (INTC -1.16%) has been trying to gain ground. The company's Gaudi line of AI chips historically couldn't match Nvidia on performance, but Intel pitched them as a cost-effective alternative to Nvidia's pricey graphics processing units (GPUs).
Gaudi 3, Intel's third-generation Gaudi AI chip, was officially unveiled on Tuesday. This time around, Intel is pushing both the performance and cost angles. The company claims that Gaudi 3 beats Nvidia's H100 GPU by 50%, on average, in AI inference tasks, with 40% better power efficiency -- and it will be available at "a fraction of the cost" of Nvidia's powerful chip.
Massive performance gains
Intel presented impressive numbers, although they should be taken with a grain of salt until third-party benchmarks are performed. The company claims that Gaudi 3 will offer quadruple the AI compute performance over Gaudi 2 using a particular low-precision format and 50% more memory bandwidth.
On top of beating Nvidia's H100 in AI inference, Intel also claims that Gaudi 3 outperforms the H100 in multiple AI training tasks. Across three different large language models, Intel claims that Gaudi 3 completes training 50% faster. How these figures change when using a large cluster of Gaudi 3 chips versus a large cluster of H100 chips is unclear, since performance will depend on data-transfer speeds and other factors.
Building an open ecosystem
While Nvidia's AI chips are extremely powerful, the company's proprietary software is partly responsible for its dominance. Nvidia's CUDA platform has become the industry standard for accelerated computing and only works with Nvidia hardware.
Intel and other tech giants are looking to change this situation. Intel technology is at the center of the Unified Acceleration Foundation, a consortium of companies working on an open alternative to Nvidia's CUDA. Separately, Intel is working with companies like Hugging Face, RedHat, Redis, SAP, and VMWare to create an open platform for enterprise AI. Intel's Gaudi 3 also works with industry-standard Ethernet networking, rather than proprietary solutions.
Intel's Gaudi 3 will be available to original equipment manufacturers (OEMs) building AI servers in the second quarter of 2024, with general availability expected in the third quarter. Intel has Dell, HPE, Lenovo, and Supermicro on board as partners.
Opportunity on multiple fronts
Estimates for how big the AI chip market will become over the next decade vary, but it's safe to say that demand for AI accelerators is set to grow. Intel can benefit in two ways.
First, its Gaudi family of chips and future AI accelerators can go after the AI chip market directly. As software libraries and applications are ported to work with a wide range of AI hardware, Intel's focus on undercutting Nvidia on price while delivering solid performance and efficiency should bring in plenty of customers.
Second, Intel is aiming to become the world's second-largest foundry by 2030. Gaudi 3 is manufactured by TSMC, but as Intel ramps up production on its advanced process nodes, it may eventually move manufacturing for its AI chips in-house. Other companies looking to design their own AI chips may also turn to Intel for manufacturing or packaging services, allowing Intel to capture additional revenue from the AI chip market.
While time will tell whether Intel's Gaudi 3 will win meaningful market share from Nvidia, the long-term picture looks bright for the chip giant as demand for AI chips explodes.
