Is there a theoritical minimum for computing power required to say, target GPT-2? Is there something fundamental to prevent a gaming laptop from exceeding Claude Opus?
If you'd do this on a gaming laptop, it'd all be on SSDs, which are orders of magnitude slower than GPU's for memory access
Also, AI uses maths, called FLOPS, floating point operations
My laptop cpu (7840U) has 4.1TFLOPS, a H200 GPU has 3,958 TFLOPS
OpenAI chatgpt 5 was reportedly trained on ~100-200k nvidia GPU's
So:
- accessing data is 1000x slower
- maths is 1000x slower
- they have up to 200,000x more GPU's than a laptop
Now remember each part of the data is used multiple times, you start getting into the GPU's being 1000x1000x200,000x( data access multiple times) faster
So, I don't think there's fundamentally something impossible with training claude opus on your laptop, but moreso the time required would be so infinitely high that it's very improbable.
TylerLives 2 days ago [-]
You could do it by hand, by calculating the gradients and doing backprop with pen and paper.
JPLeRouzic 2 days ago [-]
I have an unrelated question: Why the URL is submission is "twitter.com" when the link leads to "x.com". Is it still possible to use twitter.com?
GuestFAUniverse 2 days ago [-]
And why becomes "~" (circa) a minus?
Luckily that error is blatantly obvious.
clawsyndicate 2 days ago [-]
[dead]
Rendered at 18:03:38 GMT+0000 (Coordinated Universal Time) with Vercel.
Bear in mind this is: - 5 years old - only cpu
If you'd do this on a gaming laptop, it'd all be on SSDs, which are orders of magnitude slower than GPU's for memory access
Also, AI uses maths, called FLOPS, floating point operations
My laptop cpu (7840U) has 4.1TFLOPS, a H200 GPU has 3,958 TFLOPS
OpenAI chatgpt 5 was reportedly trained on ~100-200k nvidia GPU's
So: - accessing data is 1000x slower - maths is 1000x slower - they have up to 200,000x more GPU's than a laptop
Now remember each part of the data is used multiple times, you start getting into the GPU's being 1000x1000x200,000x( data access multiple times) faster
So, I don't think there's fundamentally something impossible with training claude opus on your laptop, but moreso the time required would be so infinitely high that it's very improbable.
Luckily that error is blatantly obvious.