Unverified50% confidenceFactExact time
典型的张量并行度通常为2、4或8，受限于单机NVLink带宽，跨节点张量并行代价极高
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
非均匀张量并行：NVIDIA提升LLM训练Goodput的容错新方案
rss7/6/2026
Related Claims
UnverifiedThe fifth generation of NVLink (featured in the Blackwell architecture) offers a single-link bidirectional bandwidth of up to 1.8 TB/s.71% similarUnverifiedIn an 8-card H20 configuration, the tensor parallelism degree is typically set to 8 to maximize VRAM utilization and keep inter-card communication overhead within the NVLink bandwidth range.69% similarVerifiedNVLink 4.0的双向带宽高达900GB/s，而传统PCIe总线带宽约为64GB/s69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/197028API
curl https://kongchang.com/api/v1/knowledge/claims/197028MCP
get_claim(id=197028)