Unverified50% confidenceTradeoffExact time

张量并行通常只在节点内使用，流水线并行则更适合跨节点场景，因NVLink为微秒级延迟而跨节点InfiniBand为毫秒级

1

Sources

50%

Confidence

Long-term

Relevance

7/11/2026

First Seen

Sources

TensorRT多设备推理：突破单GPU显存瓶颈的工程实践

rss6/25/2026

Related Claims

UnverifiedAscend C vector programming uses a double-buffering mechanism; through TPipe, a double-buffer pipeline can be conveniently constructed, and setting the queue depth to 2 automatically enables double buffering.59% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/485165

API

curl https://kongchang.com/api/v1/knowledge/claims/485165

MCP

get_claim(id=485165)