Intel has officially announced the launch of its Gaudi 3 AI accelerator, positioning it as a cost-effective alternative to NVIDIA's dominant H100 GPU in the rapidly expanding artificial intelligence market.
The Gaudi 3 represents a significant leap forward from its predecessor, boasting 64 tensor processor cores (TPCs) and 8 matrix multiplication engines (MMEs). Each accelerator is equipped with 128GB of high-bandwidth memory (HBM2E) and delivers up to 1,856 TFLOPS of FP8 performance. Intel has streamlined the architecture to focus on FP8 and BFloat16 operations, which are crucial for modern AI workloads.
While the raw performance metrics of Gaudi 3 fall short of NVIDIA's H100 in some areas, Intel is banking on its pricing strategy to gain market share. The company claims that Gaudi 3 offers up to 80% better performance per dollar compared to the H100 for certain AI models. Intel has indicated that an eight-Gaudi 3 accelerator kit will be priced at approximately $125,000, translating to about $15,625 per unit - a stark contrast to the current $30,000+ price tag of a single NVIDIA H100.
Key features of the Gaudi 3 include:
- 64 Tensor Processing Cores (TPCs)
- 8 Matrix Multiplication Engines (MMEs)
- 128GB HBM2E memory with 3.67 TB/s bandwidth
- 96MB on-die SRAM cache
- 24 200GbE networking interfaces
- 600W TDP
Intel is touting significant performance gains in AI inference tasks. For the LLaMA 3 8B model, Gaudi 3 purportedly offers a 9% improvement in inference speed while delivering 80% better performance per dollar versus the H100. With larger models like LLaMA 70B, Intel claims a 19% better inference throughput and twice the performance per dollar.
The Gaudi 3 will be available in various form factors, including OAM-compliant cards and PCIe add-in cards. Intel has partnered with major server manufacturers such as Dell, HPE, and Supermicro to integrate Gaudi 3 into their AI-focused systems. Cloud availability is also planned, with IBM Cloud and Intel's own Tiber Developer Cloud set to offer Gaudi 3 resources.
As the AI hardware race intensifies, Intel's strategy with Gaudi 3 appears to be focused on providing a more affordable option for organizations looking to deploy AI at scale. However, with NVIDIA's upcoming Blackwell architecture on the horizon, it remains to be seen whether Intel's price advantage will be enough to carve out a significant portion of the AI accelerator market.
The general availability of Gaudi 3-based systems is expected in Q4 2023, with some partners like Dell and Supermicro shipping as early as October. This launch marks a critical moment for Intel as it seeks to establish itself as a serious contender in the AI hardware space dominated by NVIDIA.
 |
|---|
| Intel's versatile AI solutions are designed to cater to diverse industries, enhancing technological advancements across sectors |