chips infrastructure
Zeros Pushed Ternary Weights Below 1.58 Bits
BITCOS reached 1.485 bits per weight on the sparsest tested model and improved decode throughput up to 1.27 times.
Summary
BITCOS reached 1.485 bits per weight on the sparsest tested model and improved decode throughput up to 1.27 times.
Ternary models are usually packed as if minus one, zero and plus one were equally likely. Measurements across 29 models found zeros reaching 51.5% of weights. BITCOS stores a presence bitmap plus compacted signs, costing two minus the zero density in bits per element. It beat five-trit packing in 26 of 29 tested models and reached 1.485 bits per weight on the sparsest. Optimized unpacking improved production kernel throughput up to 1.28 times and end-to-end decode up to 1.18 times on CPUs and 1.27 times on GPUs in the reported tests.
Why it matters
BITCOS reached 1.485 bits per weight on the sparsest tested model and improved decode throughput up to 1.27 times.
Limits and context
No additional limitation was separately recorded.
Key claims
BITCOS reached 1.485 bits per weight on the sparsest tested model and improved decode throughput up to 1.27 times.
Evidence: source-2026-09-16-008
Sources
- arXiv preprint 2609.16338arXiv · primary research
Corrections
No corrections have been recorded for this story.