Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

Unverified50% confidenceCautionExact time

FP8动态范围较窄，在反向传播和梯度累积时容易出现数值溢出或下溢，因此训练侧actor保持BF16精度

1

Sources

50%

Confidence

Long-term

Relevance

7/12/2026

First Seen

Sources

DeepSeek-V4 Flash RL训练成功迁移AMD MI355X GPU全解析

twitterlmsysorg7/10/2026

Related Claims

Verified混合精度训练将权重保存为FP32，而前向和反向传播计算使用FP16或BF16，可将显存占用减少近一半84% similar Unverified混合精度训练通常配合损失缩放（Loss Scaling）技术使用，以防止梯度在 FP16 表示范围内消失76% similar UnverifiedFP8与INT8整数量化不同，FP8保留了浮点数的动态范围，对模型精度影响更小73% similar Unverified模型量化从FP32降至INT8能将推理速度提升3-4倍，但可能导致小目标检测精度下降70% similar UnverifiedFP8推理加速效果依赖工作负载特征：计算密集型大批量矩阵运算可接近理论2倍加速，显存带宽密集型小批量推理收益有限68% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/490940

API

curl https://kongchang.com/api/v1/knowledge/claims/490940

MCP

get_claim(id=490940)