![Vivian Score Fang Profile](https://pbs.twimg.com/profile_images/998428076135530496/VVRa2GHb_x96.jpg)
Vivian Score Fang
@FangScore
Followers
87
Following
8K
Statuses
2K
A computer programer who thinks Raft is merely a specialized application of Paxos. And a dumb one at that.
California, USA
Joined May 2018
@MaxwellFrostFL Separation of power. They belong to the executive branch, you’re in the legislative branch. Chill
0
0
0
@DanCollins2011 Chinese government has been funded by selling land leases during real estate bubble. Now they are broke
0
0
0
@pmarca @gospaceport @wasnt__me_ Hats off to @ggerganov for impressive achievement in llama.cpp that allows super efficient inference in all kinds of devices. Looking forward to the Nvidia Digits for exactly this kind of scenario!
0
0
0
@KenrikMarch @pmarca There are always trade offs between vertical performance optimization and horizontal abstraction for flexibility. When you write everything in assembly (PTX in their case) not only you lock yourself in nvidia platform, you lose some flexibility for experimentation
0
0
0
@untitled01ipynb @pmarca Deepseek uses fp8 block quantization, it’s nothing new, just take a look at llama.cpp
1
0
0
@pmarca Yeah I still remember Huawei made “breakthrough” on their own cellphone soc, turned out it’s a hack with low volume and cost overruns. And then they still have to use shell companies to get chips from tsmc
0
0
0
@WholeMarsBlog No they did not, we know all the perf optimizations in their paper and have tried them all, some of them don’t work as well as they claimed.
0
0
1
@shrihacker If you think cuda is an nvidia lock in, ptx is an even stronger lock in. They went to ptx just showed how programmable nvidia chips are. None of the competitors chips have this kind of flexibility t
0
0
1