TechMakesArt
Region: US
Sunday 19 April 2026 14:15:15 GMT
3512
118
14
4
Music
Download
Comments
Otto Fjällström :
what if you use a m.2 to pcie adapter and a server-grade nic that supports rdma? It will get rid of most of the latency between nodes. For power savings you could underclock and undervolt the gpu. You should get as low as 150w per card and not loose much performance on MoE models.
2026-05-12 01:30:59
1
Fox :
I'm still building up my homelab, but I have a custom RCCL driver build script derived from another custom project. I'm hoping that allows getting closer, but the limitation is still the onboard gigabit networking.
2026-06-05 02:12:11
0
prometheus02 :
what about a council of smaller.models? 🤔
2026-04-19 14:28:02
3
smeeegol :
You can have the fastest embeddings server evar
2026-04-22 22:19:45
0
bird :
They're okay for 4-9B dense models and dedicated embeddings. I'm also looking at using them for disaggregated prefill.
2026-04-30 19:35:21
0
mark :
I would assume that hardware modifications are required to have probably 10 GB ethernet and a 10 gig switch so that all the nodes can communicate to each other.
2026-05-15 22:33:36
0
Gonzales :
Just sell them individually at a reasonable price please. 😌
2026-04-19 19:12:35
0
HansCCT Gaming :
Could you do FoldingAtHome on these?
2026-04-19 16:23:08
0
prometheus02 :
🥺
2026-04-19 14:27:19
0
To see more videos from user @techmakesart, please go to the Tikwm
homepage.
