Ha, friends who have AI accounts in Xiaohongshu, we must never write about Deepseek related topics.
The deepseek notes with more than 10,000 likes are all about abstraction. A technical man comes in and talks about how Deepseek Harness is uncompetitive. The soul painter has transformed deepseek into a male beauty whale. In this dimension, words can be expressions, movements, music... It’s hard to understand.
Su Zifeng recently demonstrated a host machine the size of a lunch box, running a large model with 235 billion parameters on site.
It doesn’t rely on a data center or the cloud, it relies entirely on AMD’s Ryzen AI Max+ 395, 128GB unified memory, and the CPU and GPU share the memory pool. The RTX 5090 has 32GB of video memory, the RTX 4090 has 24GB, and this small host has 128GB, which is more than three times the capacity of the flagship independent graphics.
I saw a press conference and I was dumbfounded. Su Zifeng came to the stage carrying a host computer…
I saw a press conference and I was dumbfounded. Su Zifeng came to the stage carrying a host computer the size of a lunch box, and ran a large model with 235 billion parameters on the spot. No need to use the cloud or rent a GPU, all relying on AMD Ryzen AI Max+ 395, 128GB unified memory. The RTX 5090 only has 32GB of video memory, and the 4090 has 24GB.
AMD CEO Su Zifeng demonstrated a mini host equipped with Ryzen AI Max+ 395 processor, which can locally run a large model with 235 billion parameters.
The device uses a 128GB unified memory architecture, and the CPU and GPU share the same memory pool, which is a first in the field of x86 processors. Specification comparison shows that RTX 5090 has 32GB of video memory, and RTX
You have reached the end of the archive
All of deepseek