跳到正文
The Elec· Hyun-Seon, Park·· 3 天前AI 評分38

SiMa.ai 推出 Palette 3.0,支援 Physical AI 開發

SiMa.ai Unveils Palette 3.0 for Physical AI Development

AI 導讀

SiMa.ai 推出 Palette 3.0,這個 Agentic 軟件平台面向 Physical AI 模型,支援多個模型並行執行,以處理真實環境中的推理和行動。

正文

SiMa.ai, a global physical AI technology company, has unveiled Palette 3.0, an agentic software platform designed to deliver high-performance execution across physical AI models. First introduced in June, SiMa.ai Palette is an agentic development platform that automates the development of physical AI applications on SiMa.ai's hardware platform.

According to SiMa.ai, Palette 3.0 supports AI reasoning and actions in real-world environments through enhanced concurrent execution of multiple models. The company said the platform can efficiently run multiple AI models simultaneously, including through separate Docker containers on a single Modalix Machine Learning System-on-Chip (MLSoC), SiMa.ai's dedicated edge AI processor. This enables a range of physical AI capabilities using large language models (LLMs) and vision-language models (VLMs).

The platform also improves processing performance for perception, depth estimation and reasoning models required for drones and robotics. In collaboration with partners, SiMa.ai has added integrated sensor fusion, Dense Neural Depth Fields and closed-loop digital twin simulations. The company said these capabilities can accelerate the development of autonomous systems.

SiMa.ai has also introduced new open-source capabilities for generative AI application development. The company said Palette 3.0 incorporates the latest EAGLE3, DFlash and Gemma 4 MTP speculative decoding algorithms, increasing model throughput by up to 3.5 times in tokens per second (TPS), even in low-power environments.

Large-scale mixture-of-experts (MoE) models, including OpenAI's GPT-OSS-20B, can also run on a 16GB system-on-module (SoM) powered by SiMa.ai's Modalix MLSoC.

"The fundamental bottleneck in physical AI today stems from the gap between software that evolves every month and conventional silicon hardware that takes years to adapt," SiMa.ai said. "While other companies adapt data center technologies for physical AI, SiMa.ai develops software and silicon specifically designed to meet the requirements of physical AI."

來源:The Elec · thelec.net