this post was submitted on 26 Nov 2023
1 points (100.0% liked)

LocalLLaMA

3 readers
1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 1 year ago
MODERATORS
 

https://www.amazon.se/-/en/NVIDIA-Tesla-V100-16GB-Express/dp/B076P84525 price in my country: 81000SEK or 7758,17 USD

My current setup:
NVIDIA GeForce RTX 4050 Laptop GPU
cuda cores: 2560
memory data rate 16.00 Gbps

My laptop GPU works fine for most ML and DL tasks. I am currently finetuning a GPT-2 model with some data that I scraped. And it worked surprisingly well on my current setup. So it's not like I am complaining.

I do however own a stationary PC with some old GTX 980 GPU. And was thinking of replacing that with the V100.

So my question to this community is: For those of you who have bought your own super-duper-GPU. Was it worth it. And what was your experience and realizations when you started tinkering with it?

Note: Please refrain giving me snarky comments about using Cloud GPU's. I am not interested in that (And I am in fact already using one for another ML task that doesn't involve finetuning) . I am interested to hear about the some hardware hobbyists opinion on this matter.

you are viewing a single comment's thread
view the rest of the comments
[–] ambient_temp_xeno@alien.top 1 points 11 months ago (5 children)
[–] aikitoria@alien.top 1 points 11 months ago

Is there any such benchmark that includes both the 4090/A100 and a mac with M2 Ultra / M3 Max? I've searched quite a bit but didn't find anyone comparing them on similar setups, it seems very interesting due to the large unified memory.

[–] freecodeio@alien.top 1 points 11 months ago (1 children)

So basically either 4090 or H100

[–] holistic-engine@alien.top 1 points 11 months ago

Yeah, perhaps If I am crazy enough I could just buy 3 of those and call it a day

[–] az226@alien.top 1 points 11 months ago

A6000 being worse than 3090 doesn’t make any sense.

[–] Mescallan@alien.top 1 points 11 months ago

Man those h100s really are on another level. I shudder to think where are in 5 years.

[–] FullOf_Bad_Ideas@alien.top 1 points 11 months ago (1 children)

I can't corraborate results for Pascal cards. They had very limited FP16 performance, usually 1:64 of FP32 performance. Switching over to rtx 3090 ti from gtx 1080 got me around 10-20x gains in qlora training, assuming keeping the exact same batch size and ctx length, changing only calculations from fp16 to bf16.

[–] ambient_temp_xeno@alien.top 1 points 11 months ago

I'm not sure where this chart is from, but I remember it was made before qlora even existed.