Intel Gaudi 3 AI Accelerators To Cost Around Half Of NVIDIA’s Blackwell Pricing

Introduction

At Computex 2024, Intel made waves by unveiling its Gaudi 3 AI accelerators. These chips are designed to tackle the generative AI gap, and they come with an enticing price tag that’s sure to turn heads. Let’s break down the details.

The Pricing Surprise

Intel’s Gaudi 3 accelerator kit consists of eight individual units. The entire kit is priced at $125,000, which means each Gaudi 3 AI accelerator costs approximately $15,625. In comparison, the older Gaudi 2 AI accelerator kit, with the same number of units, is priced at around $8,125. But here’s the real shocker: Intel’s Gaudi 3 is roughly half the cost of competing GPUs from Nvidia, such as the Blackwell B100 AI GPU, which is expected to retail at a hefty $35,000 per unit.

Performance Per Dollar

While affordability is a key selling point, it’s essential to consider the bigger picture. Nvidia’s ecosystem, including specialized CUDA libraries and Grace CPUs, has contributed to its widespread adoption in the AI market. Intel, on the other hand, aims to win over buyers not just with lower prices but also by emphasizing performance improvements.

• Training Speed: An 8,192-accelerator cluster of Gaudi 3 claims up to 40% faster time-to-train compared to an equivalent Nvidia H100 GPU cluster.
• Training Throughput: For a 64-accelerator cluster, Gaudi 3 delivers up to 15% faster training throughput on specific models.
• Inferencing Speed: Gaudi 3 achieves up to 2x faster inferencing speed on popular large language models (LLMs) compared to Nvidia H100.

The Battle Ahead

Intel faces an uphill battle against Nvidia’s dominance. While Gaudi 3’s affordability is compelling, Nvidia’s reputation for superior products remains strong. Buyers must weigh raw power against ecosystem advantages. Can Intel’s Gaudi 3 disrupt the status quo? Only time will tell, but the AI race just got more interesting! 💡

Stay tuned for further developments in the world of AI accelerators. Until next time, tech enthusiasts! 👾