Software Engineering Courses (SECourses)•10mo ago

FP8 will be extra slow on any gpu before the nvidia Lovelace generation like A6000, 3090, A100 dont

FP8 will be extra slow on any gpu before the nvidia Lovelace generation
like A6000, 3090, A100 dont have the chips needed to perform well with FP8

L40, 4090, h100 would be good cards for FP8

just incase anyone cares

FFurkan Gözükara SECourses it is quality vs time

Jean-lu•2/26/25, 3:37 PM

Any idea how fal does it? the speed is amazing & quality is almost perfect too. Are they finetuning layers or just good GPUs

Ccomputersheep FP8 will be extra slow on any gpu before the nvidia Lovelace generation like A6...

Furkan Gözükara SECourses•2/26/25, 3:44 PM

nice info thanks

JJean-lu Any idea how fal does it? the speed is amazing & quality is almost perfect too. ...

Furkan Gözükara SECourses•2/26/25, 3:44 PM

yes i know i made a research for a client

Furkan Gözükara SECourses•2/26/25, 3:44 PM

higher batch size bigger GPU multiple GPU

Furkan Gözükara SECourses•2/26/25, 3:44 PM

lower steps higher LR

Furkan Gözükara SECourses•2/26/25, 3:44 PM

JJean-lu Any idea how fal does it? the speed is amazing & quality is almost perfect too. ...

Timson•2/26/25, 4:00 PM

increase learning rate until you are no longer satisfied with quality

Timson•2/26/25, 4:02 PM

lowering resolution is also an option

Timson•2/26/25, 4:02 PM

also decreasing dataset diversity will speed up likeness gain at the cost of generalization

Zaulinator•2/26/25, 4:10 PM

14b i2v on default settings with 24gb Vram preset is almost 1hr30 to generate the video, oooof

Furkan Gözükara SECourses•2/26/25, 4:15 PM

ZZaulinator 14b i2v on default settings with 24gb Vram preset is almost 1hr30 to generate th...

Furkan Gözükara SECourses•2/26/25, 4:16 PM

yep true

Furkan Gözükara SECourses•2/26/25, 4:16 PM

it takes 3 hours on 3090 ti

Furkan Gözükara SECourses•2/26/25, 4:16 PM

i hope this works time to test 5090

Zaulinator•2/26/25, 4:20 PM

1.3b txt2video is actually far faster than i would have thought

Zaulinator•2/26/25, 4:21 PM

under 2 mins

ZZaulinator 1.3b txt2video is actually far faster than i would have thought

Furkan Gözükara SECourses•2/26/25, 4:22 PM

yep super fast

Furkan Gözükara SECourses•2/26/25, 4:22 PM

and super high quality

Furkan Gözükara SECourses•2/26/25, 4:24 PM

yes flash attention compiled

Furkan Gözükara SECourses•2/26/25, 4:24 PM

now gonna test

TheToday•2/26/25, 4:35 PM

Is it possible to create lora in wan?

TTheToday Is it possible to create lora in wan?

Furkan Gözükara SECourses•2/26/25, 5:06 PM

not atm i think

Furkan Gözükara SECourses•2/26/25, 5:06 PM

rtx 5090 speed native python 3.10 venv

Furkan Gözükara SECourses•2/26/25, 5:07 PM

awesome 5090 workng

Furkan Gözükara SECourses•2/26/25, 5:07 PM

amazing

Furkan Gözükara SECourses•2/26/25, 5:07 PM

now i can update all of our apps

1nfern0•2/26/25, 5:07 PM

Thats not terrible..

Furkan Gözükara SECourses•2/26/25, 5:07 PM

let me test fp8 if will work

Furkan Gözükara SECourses•2/26/25, 5:07 PM

this is bf16

Furkan Gözükara SECourses•2/26/25, 5:08 PM

522 watt pretty good utilization

Furkan Gözükara SECourses•2/26/25, 5:14 PM

RTX 5090 vs 3090 TI

1nfern0•2/26/25, 5:21 PM

Nice..

Furkan Gözükara SECourses•2/26/25, 5:22 PM

yep

Furkan Gözükara SECourses•2/26/25, 5:22 PM

and this is not max perf

Furkan Gözükara SECourses•2/26/25, 5:22 PM

lots of other apps

Furkan Gözükara SECourses•2/26/25, 5:22 PM

now i can make every app work

11nfern0 Nice..

sergio_gonzo•2/26/25, 5:36 PM

Dr Furkan the app that you shared have the image to video? but can work in a 4070 super ti 16 gigas vram?

Ssergio_gonzo Dr Furkan the app that you shared have the image to video? but can work in a 407...

Furkan Gözükara SECourses•2/26/25, 5:48 PM

yes but slow

Furkan Gözükara SECourses•2/26/25, 5:48 PM

so reduce resolution

Zaulinator•2/26/25, 6:44 PM

@Dr. Furkan Gözükara - what settings are currently giving you the cleanest output for txt2video?

FFurkan Gözükara SECourses now i can make every app work

1nfern0•2/26/25, 6:58 PM

My poor Drive

1nfern0•2/26/25, 6:58 PM

So for the 5090, should I use the 24Gb profile?

1nfern0•2/26/25, 7:20 PM

Look forward to some smaller models for the 14b. This is on the 5090

11nfern0 Look forward to some smaller models for the 14b. This is on the 5090

Zaulinator•2/26/25, 7:23 PM

I cant get either 14b model to run on my 4090. tried every vram preset

ZZaulinator @Dr. Furkan Gözükara - what settings are currently giving you the cleanest outp...

Furkan Gözükara SECourses•2/26/25, 7:38 PM

default

11nfern0 My poor Drive

Furkan Gözükara SECourses•2/26/25, 7:38 PM

i am gonna add 32 gb config

Furkan Gözükara SECourses•2/26/25, 7:38 PM

that will use 30 gb

11nfern0 Look forward to some smaller models for the 14b. This is on the 5090

Furkan Gözükara SECourses•2/26/25, 7:38 PM

ye a 7b model would be amazing

Furkan Gözükara SECourses•2/26/25, 7:39 PM

even 1.3b strong

ZZaulinator I cant get either 14b model to run on my 4090. tried every vram preset

Furkan Gözükara SECourses•2/26/25, 7:39 PM

if you have ram should work

FP8 will be extra slow on any gpu before the nvidia Lovelace generation like A6000, 3090, A100 dont

Similar Threads