Artificial Intelligence thread
- Thread starter 9dashline
- Start date
The PLA military procurement's philosophy is '探索一代、预研一代、研制一代、生产一代‘ - 'explore a generation, research a generation, develope a generation, produce a generation' - I imagine there would be similar ethos in Chinese AI companies. I would be very surprised if they don't have something that's better up their sleeves.
So for tokens per second:
Cerebras ??: 1600
Nvidia H200: 3800
I don't think Steve hsu knows what he's talking about. Cerebras, Groq, Sambanova, etc are gimmicks.
Cerebras probably claimed the 57x number off of this.
Indeed on that page you funnily enough see that 57x number come up again, that being the WSE-3 is 57x bigger than a H100's die...
Last edited:
People want to use it. Western media can talk shit about censoring and flawed security but since it's open source all cloud providers need to host it unless they want to give money and market share to their competitors. As soon as Microsoft's Azure made the jump rest will all have to follow.
Username checks out.
Cerebras is not running the full version. They are running the 70b distilled version, which you can run on many other lower powered chips.
Nvidia is running the full 671B.
"UC Berkeley researchers have developed a small-scale language model reproduction of DeepSeek R1-Zero, an AI language model developed in China, for about $30.[...]
The language model TinyZero is a project led by campus graduate researcher Jiayi Pan and three other researchers, advised by campus professor Professor Alane Suhr and University of Illinois at Urbana-Champaign assistant professor Hao Peng.[...]TinyZero is a small-scale reproduction, with the $30 price going toward server costs to run the experiments. TinyZero is “only useful for very restricted types of tasks” such as countdown and multiplication tasks[...]"
The language model TinyZero is a project led by campus graduate researcher Jiayi Pan and three other researchers, advised by campus professor Professor Alane Suhr and University of Illinois at Urbana-Champaign assistant professor Hao Peng.[...]TinyZero is a small-scale reproduction, with the $30 price going toward server costs to run the experiments. TinyZero is “only useful for very restricted types of tasks” such as countdown and multiplication tasks[...]"