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/in/David& 39;s random thoughts/com.david_random/566

2.4K viewsApr 24 at 10:42

tg-me.com/david_random/566

Create: 2025-04-24
Last Update: 2025-06-27 04:08:22

BY David's random thoughts

Share with your friend now:
tg-me.com/david_random/566

David& 39;s random thoughts Telegram | DID YOU KNOW?

What is Telegram Possible Future Strategies?

Cryptoassets enthusiasts use this application for their trade activities, and they may make donations for this cause.If somehow Telegram do run out of money to sustain themselves they will probably introduce some features that will not hinder the rudimentary principle of Telegram but provide users with enhanced and enriched experience. This could be similar to features where characters can be customized in a game which directly do not affect the in-game strategies but add to the experience.