Hacker News
new
|
past
|
comments
|
ask
|
show
|
jobs
|
submit
login
radq
on March 30, 2024
|
parent
|
context
|
favorite
| on:
Qwen1.5-Moe: Matching 7B Model Performance with 1/...
1/3rd "activated parameters", while also requiring 2x the VRAM.
YetAnotherNick
on March 30, 2024
|
next
[–]
That's the point of MoE. Sacrificing VRAM for compute/RAM bandwidth which makes it harder sell for consumer devices but easier for server devices where things are more likely to be compute or RAM bandwidth bound.
Guidelines
|
FAQ
|
Lists
|
API
|
Security
|
Legal
|
Apply to YC
|
Contact
Search: