Tenstorrent Unveils Galaxy AI Platform Targeting Scale and Efficiency (www.forbes.com)

Tenstorrent has launched its Galaxy AI platform, featuring the Blackhole system designed to enhance AI infrastructure's efficiency and performance for real-world applications. Unlike existing systems that focus primarily on peak compute throughput, Galaxy emphasizes sustained inference performance, high-speed memory access, and scalable networking, all critical as AI workloads increase in size and concurrency. This architectural shift is pivotal for the AI/ML community as it demonstrates a growing recognition that effective AI deployments hinge on data movement efficiency rather than sheer compute power alone. The Galaxy system integrates 32 Blackhole ASICs based on RISC-V architecture, providing an impressive 23 PFLOPS of Block FP8 AI compute. However, the standout features are its memory capabilities, including 6.2 GB of on-chip SRAM and 1 TB of external GDDR6 memory, yielding high bandwidth and reducing data movement latency—a crucial factor as model sizes grow. Additionally, Tenstorrent's focus on Ethernet-based networking rather than proprietary interconnects facilitates better scalability across distributed clusters, which is essential for large-scale AI applications. This innovative approach aims to enhance resource utilization and service reliability, fostering a new standard for AI infrastructure as it evolves.