chips infrastructure
Compression Targets the Hardware, Not the Parameter Count
AutoCompress searches architectures against latency, throughput and memory models validated on actual serving stacks.
Summary
AutoCompress searches architectures against latency, throughput and memory models validated on actual serving stacks.
The authors report higher average accuracy than Minitron across RTX 5070, RTX 5080 and H100 settings, and compress Nemotron 3.5 Super from 120 billion to 74 billion parameters for one B200 at 128K context while retaining 95 percent of its mean benchmark score.
Why it matters
AutoCompress searches architectures against latency, throughput and memory models validated on actual serving stacks.
Limits and context
No additional limitation was separately recorded.
Key claims
AutoCompress searches architectures against latency, throughput and memory models validated on actual serving stacks.
Evidence: source-2026-11-06-017
Sources
- arXiv preprint 2611.08290, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.