Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
frizdny5
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
4 ms
·
1.
▲
by
frizdny5
2y ago
The bottleneck for LLM is fast and large memory, not compute power. Whoever is recommending investing in better chip(ALU) design hasn't done even a basic analysis of the problem. Tokens per second = memory bandwidth divided by model si