Quadric Chimera GPNPU Supports Variable-Length Prompts
Understand the story
According to a November 11, 2026 report by Semiconductor Engineering, Quadric's Chimera GPNPU supports processing prompts of arbitrary lengths with a single compiled graph, ranging from short voice commands to thousands of tokens, up to the configured capacity. The report notes that traditional compiled inference stacks typically run for a fixed length (e.g., 1024 tokens), relying on padding to fill remaining positions for short prompts, and require building and validating multiple binaries for different lengths. In contrast, the Chimera GPNPU needs only a single binary to cover all prompt lengths, thereby reducing compute waste from padding and lowering multi-version maintenance costs. Currently this report is the sole source of information on the event, with no other progress reported.
AI overview of reporting · Updated 2 h前
Reporting timeline
Explore different aspects of the story through its reporting.
- Semiconductor Engineering Chips and PackagingVariable-Length Prompts on Chimera GPNPU
The Quadric Chimera GPNPU supports a single compiled graph handling any prompt length, from short voice commands to thousands of tokens, up to its configured capacity. Traditional compiled inference stacks typically target fixed lengths (e.g., 1024 tokens), with padding computations filling unused positions for short prompts, requiring multiple binaries for different lengths to be built and validated individually. Chimera GPNPU needs only one binary to cover all lengths, cutting padding waste and multi-version maintenance costs.
Story heat trend
Current heat9·Comparable peak10（11/11 17:00）·Comparable change over 24 hours –
The trend compares the same continuously observed sources, which may cover fewer sources than the current heat score. Hover or select the chart to inspect each hour; use the left and right arrow keys to navigate.