Intel has officially unveiled its latest artificial intelligence (AI) accelerator, the Gaudi 3, positioning it as a more affordable option compared to Nvidia's popular H100 GPU. While the Gaudi 3 may not match the raw performance of its competitor, Intel is banking on its lower price point and reduced total cost of ownership to attract customers in the rapidly growing AI market.
Technical Specifications
The Gaudi 3 boasts impressive hardware, featuring:
- 64 tensor processor cores
- 8 matrix multiplication engines
- 96MB of on-die SRAM cache
- 128GB of HBM2E memory
- 24 200 GbE networking interfaces
- 14 media engines supporting various video codecs
This new processor represents a substantial upgrade from its predecessor, the Gaudi 2, offering increased processing power and memory capacity.
Performance Comparison
When compared to Nvidia's H100, the Gaudi 3 demonstrates competitive performance in certain areas:
- BF16 matrix performance: 1,856 TFLOPS (vs. 1,979 TFLOPS for H100)
- FP8 matrix performance: 1,856 TFLOPS (vs. 3,958 TFLOPS for H100)
- BF16 vector performance: 28.7 TFLOPS (vs. 1,979 TFLOPS for H100)
While the Gaudi 3 lags behind in some metrics, Intel claims it offers a significant price-performance advantage over the H100.
Pricing Strategy
Intel's pricing strategy is aggressive, with an accelerator kit containing eight Gaudi 3 processors priced at $125,000, or approximately $15,625 per unit. This represents a substantial saving compared to the Nvidia H100, which is currently priced at around $30,678 per card.
Availability and Ecosystem
The Gaudi 3 will be accessible through various channels:
- IBM Cloud
- Intel Tiber Developer Cloud
- Systems from Dell, HPE, and Supermicro (available in Q4 2023)
As the AI hardware landscape continues to evolve, Intel's Gaudi 3 enters the market as a compelling option for organizations seeking to balance performance with cost-effectiveness in their AI infrastructure investments.