Tenstorrent's AI Chip: Challenging Nvidia's Dominance with Competitive Pricing and Scalability
0:32 Introduction: Jim Keller, renowned for his work at AMD, Tesla, and Intel, has developed an AI chip through his startup, Tenstorrent, aiming to rival Nvidia's GPUs.
1:12 Jim Keller's Legacy: Keller's past achievements include designing AMD's Athlon and Zen architectures, Apple's A4 and A5 processors, and Tesla's Hardware 3.0 chip, consistently focusing on power-to-performance efficiency.
3:40 Wormhole Chip: Tenstorrent's new AI chip, "Wormhole," comes in two models: N150 and N300, featuring the "TenSix" processor.
4:01 Performance and Efficiency: The N150 (160W TDP) and N300 (300W TDP) boast 262 and 466 teraflops respectively, using GDDR6 memory instead of HBM for cost-effectiveness.
4:28 Comparison with Nvidia: While Nvidia's H100 offers superior performance (16,670 teraflops), it's significantly more expensive ($30,000+). Tenstorrent's N300 provides about 30% of the H100's performance at a fraction of the cost ($1,399).
4:42 Workstation Offerings: Tenstorrent offers workstations, including a "loud box" (N300, $12,000) and a "quiet box" with liquid cooling (N300, $15,000), targeting individual users and HPC applications.
6:21 Scalability Focus: Tenstorrent emphasizes scalability, similar to Tesla's Dojo system, using Ethernet for interconnectivity and allowing for future expansion.
7:28 Open Standards: Unlike Nvidia's proprietary standards, Tenstorrent embraces open formats like Risk 5 and Ethernet, promoting broader accessibility and collaboration.
8:25 Design Philosophy: Tenstorrent simplifies individual unit calculations, connects them for scalability, addresses DRAM capacity bottlenecks with Ethernet, and prioritizes power efficiency.
10:13 Future Outlook: Tenstorrent's upcoming "Black Hole" version and expanded rack models aim to further challenge Nvidia's dominance by offering competitive performance and cost-effectiveness.