Unverified50% confidenceBenchmarkExact time

SPAD (Princeton/University of Washington) designs the chip from scratch: when HBM bandwidth is cut by 40%, prefill is only 17% slower, and when compute is cut by 50%, decode is only 22% slower, with cluster simulations showing a 19% to 41% reduction in hardware cost.

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

10/19/2026

First Seen

Valid until: 1/17/2027

Sources

One Machine Becomes Two: LLM Inference Disaggregation and the Engineering Logic Behind a 75% Throughput Gain

Related Entities

SPADmodel Princeton Universitycompany University of Washingtoncompany

Related Claims

UnverifiedAnthropic reports that Scion reduced infrastructure footprint (CPU and memory) by about 200 times, made checkout about 100 times faster (from several minutes to a few seconds), saved about 25% of CI capacity, achieved cache hit rates of 99%+ at each layer, and realized zero repository maintenance.67% similar UnverifiedHardware-accelerated decoding can reduce CPU usage for 4K decoding from 60–80% to below 5%, while significantly lowering power consumption.65% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/1070132

API

curl https://kongchang.com/api/v1/knowledge/claims/1070132

MCP

get_claim(id=1070132)