Unverified50% confidenceBenchmarkExact time
On a dual-NUMA PCIe architecture, CHARM+ achieves an average forward speedup of 1.3x and an average backward speedup of 1.7x, compared against Ring-Attention.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/5/2026
First Seen
Valid until: 2/3/2027
Sources
Related Entities
Related Claims
UnverifiedCompared with PyTorch SDPA, FlexAttention, and the SGLang attention kernel, CACPU achieves up to a 2.01x speedup.73% similarUnverifiedCPU推理速度通常为GPU速度的1/5到1/1068% similarUnverified配备NVIDIA GPU并使用CUDA加速，7B模型可以达到50-100+ token/s的推理速度66% similarUnverifiedOn a rented machine, inference speed was 1.17 to 1.51 times slower when the two GPUs were attached to different NUMA nodes.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1137959API
curl https://kongchang.com/api/v1/knowledge/claims/1137959MCP
get_claim(id=1137959)