Blaize GSP [Graph Streaming Processor] Streams AI Graphs at the Edge (www.jonpeddie.com)

Blaize has announced significant advancements with its Graph Streaming Processor (GSP), which revolutionizes AI processing by treating neural networks as graphs. The GSP’s innovative design features an on-chip hardware scheduler that enables depth-first execution across 16 cores, resulting in impressive performance of 16 TOPS at just 7 watts on Samsung's 14 nm process. This architecture minimizes the need for off-chip memory access, as it keeps intermediate data on-chip, leading to lower latency and higher efficiency. The upcoming GSP2 targets an ambitious 1,600 TOPS on a 5 nm process, specifically designed for automotive applications like advanced driver-assistance systems (ADAS). This development is pivotal for the AI/ML community as it highlights a paradigm shift towards graph-native execution models, which can address the memory bandwidth limitations faced by traditional AI accelerators. Blaize’s approach not only optimizes the processing of complex neural networks but also simplifies the development process for engineers through its comprehensive software stack, which includes the no-code Blaize AI Studio and the more advanced Picasso SDK. However, while the GSP architecture showcases promising performance gains, Blaize must navigate financial challenges, such as a significant operating loss and a cautious production timeline for GSP2, to ensure its long-term viability as a silicon partner in the automotive sector.