I'm fiddling around with image detection, and I've been using Colab to
build custom models. The free version is very slow, and there are other
annoyances, like it disconnects if left idle.

I'm thinking of building a linux machine to do this work, but I am
unclear how this would perform compared to colab.

If I am going to throw a bit of money at the issue, would it be better
to pay the £10 a month for the pro version of colab, rather than
spending hundreds on a machine of my own?

GB <NOTsomeone@microsoft.invalid> wrote:
> I'm fiddling around with image detection, and I've been using Colab to
> build custom models. The free version is very slow, and there are other
> annoyances, like it disconnects if left idle.
>
> I'm thinking of building a linux machine to do this work, but I am
> unclear how this would perform compared to colab.
>
> If I am going to throw a bit of money at the issue, would it be better
> to pay the £10 a month for the pro version of colab, rather than
> spending hundreds on a machine of my own?

How long do you plan to use it for? ie is this a tenner a month for a
couple of months until you get bored, or for a decade?

Not familiar with Colab specifically, but I'd guess any machine with enough
RAM would suffice, as long as it has a suitable GPU. The answer to 'what's
suitable' is almost always 'nvidia' (due to GPGPU software typically
supporting CUDA as the primary language, which is nvidia proprietary) and
then pick whatever size suits your budget. Check how much GPU RAM the
model needs, as they typically won't run on GPUs that don't have enough.

Worth looking at GPUs from generations back (1000, 1600, 2000, 3000 series)
as you may get better bang for buck with say a 1080 than a 4060.

Has anyone published benchmarks on different hardware for the models you're
using so you can compare?

If you have any desktop PCs lying around that aren't ancient (last 10
years say) you could start by putting Linux on one of those and then adding
a recent-ish (last 5 years) GPU?

Theo

On 04/10/2023 16:40, Theo wrote:
> GB <NOTsomeone@microsoft.invalid> wrote:
>> I'm fiddling around with image detection, and I've been using Colab to
>> build custom models. The free version is very slow, and there are other
>> annoyances, like it disconnects if left idle.
>>
>> I'm thinking of building a linux machine to do this work, but I am
>> unclear how this would perform compared to colab.
>>
>> If I am going to throw a bit of money at the issue, would it be better
>> to pay the £10 a month for the pro version of colab, rather than
>> spending hundreds on a machine of my own?
>
> How long do you plan to use it for? ie is this a tenner a month for a
> couple of months until you get bored, or for a decade?
>
> Not familiar with Colab specifically, but I'd guess any machine with enough
> RAM would suffice, as long as it has a suitable GPU. The answer to 'what's
> suitable' is almost always 'nvidia' (due to GPGPU software typically
> supporting CUDA as the primary language, which is nvidia proprietary) and
> then pick whatever size suits your budget. Check how much GPU RAM the
> model needs, as they typically won't run on GPUs that don't have enough.
>
> Worth looking at GPUs from generations back (1000, 1600, 2000, 3000 series)
> as you may get better bang for buck with say a 1080 than a 4060.
>
> Has anyone published benchmarks on different hardware for the models you're
> using so you can compare?

Yes, I'd spotted that NVidia is the way to go. It looks like a PC with
say a 3060 GPU is several times faster than the free Colab version.

I have a PSU and an old case lying around, but probably nothing else of
any use.

GB <NOTsomeone@microsoft.invalid> wrote:
> On 04/10/2023 16:40, Theo wrote:
>> GB <NOTsomeone@microsoft.invalid> wrote:
>>> I'm fiddling around with image detection, and I've been using Colab to
>>> build custom models. The free version is very slow, and there are other
>>> annoyances, like it disconnects if left idle.
>>>
>>> I'm thinking of building a linux machine to do this work, but I am
>>> unclear how this would perform compared to colab.
>>>
>>> If I am going to throw a bit of money at the issue, would it be better
>>> to pay the £10 a month for the pro version of colab, rather than
>>> spending hundreds on a machine of my own?
>>
>> How long do you plan to use it for? ie is this a tenner a month for a
>> couple of months until you get bored, or for a decade?
>>
>> Not familiar with Colab specifically, but I'd guess any machine with enough
>> RAM would suffice, as long as it has a suitable GPU. The answer to 'what's
>> suitable' is almost always 'nvidia' (due to GPGPU software typically
>> supporting CUDA as the primary language, which is nvidia proprietary) and
>> then pick whatever size suits your budget. Check how much GPU RAM the
>> model needs, as they typically won't run on GPUs that don't have enough.
>>
>> Worth looking at GPUs from generations back (1000, 1600, 2000, 3000 series)
>> as you may get better bang for buck with say a 1080 than a 4060.
>>
>> Has anyone published benchmarks on different hardware for the models you're
>> using so you can compare?
>
> Yes, I'd spotted that NVidia is the way to go. It looks like a PC with
> say a 3060 GPU is several times faster than the free Colab version.
>
> I have a PSU and an old case lying around, but probably nothing else of
> any use.

A 3060 is about 300 quid. Are you likely to be doing this work for more
than 2.5 years?

Your local machine may not be as fast as the google collab environment, so
you'll also be paying for ~250w of usage for however long your analysis
needs. Depending on size and complexity it could be days or weeks.

Chris <ithinkiam@gmail.com> wrote:
> A 3060 is about 300 quid. Are you likely to be doing this work for more
> than 2.5 years?
>
> Your local machine may not be as fast as the google collab environment, so
> you'll also be paying for ~250w of usage for however long your analysis
> needs. Depending on size and complexity it could be days or weeks.

Another option is to rent a cloud machine - eg AWS will sell you access to
GPU machines by the hour. Don't know how the costs stack up compared with
Google's paid version (may not be cheaper) but something to consider. It's
handy if you just want to run a quick job and not pay for having it for a
long period.

Theo