Unverified85% confidenceFactTime unknown
Groq自研的LPU（Language Processing Unit）通过将模型权重静态编译到芯片的SRAM中，消除了动态内存访问的延迟，实现数百TPS的极高吞吐量
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedGroq采用LPU架构，通过编译时静态调度实现确定性计算调度，消除运行时调度开销以获得极低且一致的延迟78% similarUnverifiedGroq开发了名为LPU（语言处理单元）的专用芯片，在某些基准测试中推理速度是传统GPU方案的10倍以上75% similarUnverifiedGroq的LPU采用TSP（Tensor Streaming Processor）架构，通过编译期确定的数据流动消除运行时动态调度开销75% similarUnverifiedGroq的LPU和Cerebras的WSE是专为推理优化的芯片，通过定制化内存架构突破带宽瓶颈73% similarUnverifiedGroq和Cerebras等专用推理芯片通过将计算单元与内存紧密集成、提升内存带宽密度来实现数十倍速度提升69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18522API
curl https://kongchang.com/api/v1/knowledge/claims/18522MCP
get_claim(id=18522)