SiMa.ai Adds Shared Chip Access in Palette 3.0, Tests 3.49x Gain
A SiMa.ai Docker example runs six applications side by side. In a Qwen3.5 9B test, decoding reached 41.75 tokens per second at about 7.5 watts.

SiMa.ai released Palette SDK 3.0 on Oct. 31, enabling separate applications and Docker containers to share the accelerator and video engines on a Modalix chip. The change allows multiple applications to run on the same chip at once.
A SiMa.ai Docker example runs six applications side by side. The release also supports concurrent language and vision processing.
In a SiMa.ai test using Qwen3.5 9B and DFlash N16, decoding reached 41.75 tokens per second, up from 11.98 tokens per second in the comparison run. That is a 3.49-fold increase. Mean system-on-chip power was approximately 7.5 watts; the result applies to that test configuration.
SiMa.ai says GPT-OSS-20B runs on a Modalix system with 16GB of memory.