Intel Challenges Nvidia Dominance with Powerful Gaudi 3 AI Chip

Intel has unveiled a new AI accelerator called Gaudi 3, which it claims is 1.7x faster than the Nvidia H100 in training AI models, and has a much lower price. The team blue also claims that Gaudi 3 is 50% better in inference workloads and consumes 40% less power than the H100.

Nvidia’s dominance in the AI chip and software market has been going on for a long time. However, Intel and AMD are working hard to develop new chips to become better competitors to Nvidia and take a share of the market. Gaudi 3 represents Intel’s most serious effort to achieve this important goal.

The third generation of the powerful Gaudi series processors is scheduled to go into mass production in Q3 2024. The processor will be produced in two form factors and will be equipped with 128 GB of HBM2e memory with a bandwidth of 3.7 terabytes per second. Intel’s new processor has 24 200 Gbps Ethernet RDMA NICs, has a power consumption of 900 watts, and delivers 1,835 TFLOPS of performance in FP8 format.

According to Intel, the performance of Gaudi 3 in FP8 and BF16 formats is two and four times higher than the previous generation, respectively. Another major generational improvement is the doubling of network bandwidth and 1.5 times the memory bandwidth.

Dell, Lenovo, and Supermicro will make Gaudi 3-powered systems available to businesses. Models of Gaudi with air cooling have been produced and production of systems with liquid cooling will begin soon.

The Gaudi 3 processor uses the same architecture as the previous generation, but is based on TSMC’s advanced 5nm lithography, while the Gaudi 2 uses TSMC’s 7nm lithography.

One of the important features for AI processors is their scalability. The more scalable they are, the more powerful data centers can be built. Intel’s Gaudi processors take a different approach than Nvidia’s B200 NVL72 systems and use 200 Gbps Ethernet connectivity.

Thanks to the doubling of network bandwidth, the Gaudi 3 processor can form clusters with up to 1,024 nodes (servers) and produce significant processing power. Each server consists of 8 Gaudi 3 processors that are interconnected.

At today’s event, Intel compared the new Gaudi 3 processor to the H100, Nvidia’s previous-generation flagship, while the green team recently unveiled the B200 processor. Intel’s slides also did not compare the Gaudi 3 to the AMD Instinct MI300.

One of the important parts of Intel’s new slides is related to power consumption, where the blue team claims that its new chip consumes 2.6% less power than its competitor in inference workloads.

Back to top button