随手摸下B200 (148SM)。Vector性能相比H100的提升比较常规(+25%)。全局原子CAS延迟是H100的2x，达到200ns，这大概就是MCM的代价。有意思的点在于wmma发生明显倒退，每SM性能回到了图形卡水平。本代大概只有tcgen05.mma可以跑出纸面tensor吞吐，这标志着计算卡重心逐渐转向类似NPU的TMEM？

David's random thoughts

随手摸下B200 (148SM)。Vector性能相比H100的提升比较常规(+25%)。全局原子CAS延迟是H100的2x，达到200ns，这大概就是MCM的代价。

有意思的点在于wmma发生明显倒退，每SM性能回到了图形卡水平。本代大概只有tcgen05.mma可以跑出纸面tensor吞吐，这标志着计算卡重心逐渐转向类似NPU的TMEM？

www.tg-me.com/hk/telegram/com.david_random/567

2.2K viewsApr 24 at 10:42

tg-me.com/david_random/567

Create: 2025-04-24
Last Update: 2025-06-01 21:12:23

BY David's random thoughts

Share with your friend now:
tg-me.com/david_random/567

telegram Telegram | DID YOU KNOW?

Date: 2025-06-01| telegram

The messaging service and social-media platform owes creditors roughly $700 million by the end of April, according to people briefed on the company’s plans and loan documents viewed by The Wall Street Journal. At the same time, Telegram Group Inc. must cover rising equipment and bandwidth expenses because of its rapid growth, despite going years without attempting to generate revenue.

Telegram is riding high, adding tens of million of users this year. Now the bill is coming due.Telegram is one of the few significant social-media challengers to Facebook Inc., FB -1.90% on a trajectory toward one billion users active each month by the end of 2022, up from roughly 550 million today.