Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Writer somewhat naive. His Ascend team couldn't get comparable performance (gen1 910A NPUs) initially vs (I assume) Nvidia because obviously. Management supported teams that pivot to cloned alternatives that used GPUs that can be immediately commercialized. Internal office politics make this happen. Ascend team works out kinks (this is huge confirmation), but feel (are) mistreated, i.e. biased bureaucracy, lack of recognition. Many burnout / leave to other Chinese AI companies.

HW strategy/culture has been burning tier1 talent since forever. I remember in the 90s When HW and other domestic PRC telco started poaching from Nortel, Siemens, Lucent etc... the talent (most Chinese diaspora used to comfy western office culture) did not have a good time fitting into an actual Chinese company with Chinese culture (but got paid lots). Many burned out too... yet HW, a particularly extreme outlier of militant work culture, has become dominant..

LBH, both HW post sanctions, is a strategic company, overlapping with semi fabrication, domestic chips, and AI is cubing their strategic value. They can get away with doing anything under the current geopolitical environment to stay dominant. The worthwhile take away from this farewell letter is HW threw enough talent at Ascend that it kind of works now, and potentially can throw enough talent at it to be competitive with Nvidia. AKA how it has always operated, like massive wankers. The intuition from the author and most of us is... you need to reward employees right, cultivate proper workplace environment blah blah blah... but look at HW for the past 30 years. They pay a lot of smart people (including patriotic suckers) A LOT of money, throw them at problems until they break. And win.



This doesn't seem right at all, given that DeepSeek reported massive performance increase because the Huawei team helped them port their LLM to HW infra. I was willing to put it in the "I don't know, maybe, let's see" category, but that comment specifically makes it read like a propaganda piece.


Ascend 910A was Q1 2022. Qwen 2.5 / Deepseek v3 was Q1 2025. Implied timeline seems to be it took ~2.5 years developement to figure out how to use Ascend somewhat competitively and HW may rival Nvidia with proper support. Where proper support in authors opinion is building a team and treating team well. My guess is Huawei can keep burning talent at problem.

The Huawei+Deepseek > Nvidia for inference claims is based on HW CloudMatrix 384 supernode using all of the optical interconnects and power to allegedly out perform Nvidia cluster. Bottleneck of Nvida cluster is switching. Bottleneck of PRC cluster is chips on old node size (slower or more power hungry). CM384 workaround is 5x more chips 910c chips and 4x more power consumption and connecting cluster with full optical to compete vs GB200 on pure cluster performance. Hard to say what actual economics of that solution is.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: