Unverified50% confidenceBenchmarkExact time
Full-loop CUDA graph capture speeds up 100-step Langevin sampling by 12.9 times (from 660 milliseconds to 51 milliseconds), and graph replay is bit-for-bit identical to eager execution.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAfter writing the KV cache in version V3, the time to generate 100 tokens from a 512-token prompt dropped from 278 seconds to 28 seconds, achieving nearly a 10x inference speedup.71% similarUnverifiedThe theoretical optimal completion time for 100 calls simultaneously hitting a token bucket (capacity 20, 10 replenished per second) is 8.0 seconds.70% similarUnverifiedAt 700,000 pixels running 10 seconds, it cannot run without tiling enabled; after enabling tiled processing and splitting into two tiles, it can scale up to 800,000 pixels and 10 seconds.70% similarUnverified在11.8万节点的生产级图基准测试中,Semantica节点搜索耗时从24毫秒降低到0.004毫秒,约6000倍加速69% similarUnverifiedImagen 4生成速度实现了10倍提升,基本10秒左右即可出图68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1099020API
curl https://kongchang.com/api/v1/knowledge/claims/1099020MCP
get_claim(id=1099020)