Microsoft is ramping up its investment in small language models (SLM), announcing the general availability of its Phi-3 models and previewing Phi-3-vision at the Build developer conference. The company also introduced Phi-3-Silica, a model specifically built for powerful neural processing units (NPUs) in devices like PC Copilot+.
Phi-3-Silica, the smallest of the Phi models with 3.3 billion parameters, will be integrated into all Copilot+ PCs on sale starting in June. Microsoft claims that the model has a latency of 650 tokens per second, using approximately 1.5 watts of power. This efficiency allows the CPU and GPU to handle other calculations, as token generation reuses the NPU’s KV cache, producing approximately 27 tokens per second when running on the CPU.
A Microsoft spokesperson highlighted Phi-Silica as Windows’ first locally implemented language model optimized for Copilot+ PC NPU, providing fast local inference. This advancement aims to empower developers to create innovative experiences for Windows users, improving productivity and accessibility within the ecosystem.
Phi-Silica is the fifth variation of Microsoft’s Phi-3 model, joining Phi-3-mini, Phi-3-small, Phi-3-medium, and Phi-3-vision, each with varying parameters. The company’s focus on SLM development reflects its commitment to enhancing AI capabilities and expanding opportunities for developers and users.
In other news, VentureBeat is hosting an exclusive event in New York on June 5 to explore strategies for auditing AI models to ensure fairness, optimal performance, and ethical compliance across different organizations. Executives are invited to attend this invitation-only event to engage with top industry leaders and delve deeper into the evolving landscape of AI auditing.
Overall, Microsoft’s emphasis on SLM development and the introduction of Phi-3-Silica for NPUs demonstrate the company’s dedication to advancing AI technology. The collaboration with industry leaders in the upcoming AI event signifies a collective effort to address biases, performance issues, and ethical challenges associated with AI models in diverse organizational settings.
Article Source
https://venturebeat.com/ai/microsoft-introduces-phi-silica-a-3-3b-parameter-model-made-for-copilot-pc-npus/