You still had a 4GB memory limit for processes, as well as a total memory limit of 64GB. Especially the first one was a problem for Java apps before AMD introduced 64bit extensions and a reason to use Sun servers for that.
You still had a 4GB memory limit for processes, as well as a total memory limit of 64GB. Especially the first one was a problem for Java apps before AMD introduced 64bit extensions and a reason to use Sun servers for that.
I was referring to work setups with the overengineering - if I had a cent for every time I had to argue with somebody at work to not make things more complex than we actually need I’d have retired a long time ago.
Unless you are gunning for a job in infrastructure you don’t need to go into kubernetes or terraform or anything like that,
Even then knowing when not to use k8s or similar things is often more valuable than having deep knowledge of those - a lot of stuff where I see k8s or similar stuff used doesn’t have the uptime requirements to warrant the complexity. If I have something that just should be up during working hours, and have reliable monitoring plus the ability to re-deploy it via ansible within 10 minutes if it goes poof maybe putting a few additional layers that can blow up in between isn’t the best idea.
Everything is deployed via ansible - including nameservices. So I already have the description of my infra in ansible, and rest is just a matter of writing scripts to pull it in a more readable form, and maybe add a few comment labels that also get extracted for easily forgettable admin URLs.
Shitty companies did it like that back then - and shitty companies still don’t properly utilize what easy tools they have available for controlled deployment nowayads. So nothing really changed, just that the amount of people (and with that, amount of morons) skyrocketed.
I had automated builds out of CVS with deployment to staging, and option to deploy to production after tests over 15 years ago.
I personally prefer bzip2 - but it needs to be packed with pbzip, not the regular bzip to generate archives that can be extracted on multiple cores. Not a good option if you have to think about Windows users, though.
Nowadays it matters if you use a compression algorithm that can utilize multiple cores for packing/unpacking larger data. For a multiple GB archive that can be the difference between “I’ll grab a coffee until this is ready” or “I’ll go for lunch and hope it is done when I come back”
As a non-Windows-user I see that as a good thing. LLMs are not going away - but that kind of nonsense at least will make sure all PCs will eventually have cheap and reasonably fast AI acceleration. Which is required for killing off centrally hosted LLMs (plus nvidias cash grabbing)
Intel is well known for requiring a new board for each new CPU generation, even if it is the same socket. AMD on the other hand is known to push stuff to its physical limits before they break compatibility.
I nowadays manage my private stuff with the ansible scripts I develop for work - so mostly my own stuff is a development environment for work, and therefore doesn’t need to be done on private time.
A lot of the Zen based APUs don’t support ECC. The next thing is if it supports registered or unregistered modules - everything up to threadripper is unregistered (though I think some of the pro parts are registered), Epycs are registered.
That makes a huge difference in how much RAM you can add, and how much you pay for it.
Not just that - intel did dual core CPUs as a response to AMD doing just that, by gluing two cores together. Which is pretty funny when you look at intels 2017 campaign of discrediting ryzen by calling it a glued together CPU.
AMDs Opteron was wiping the floor with intel stuff for years - but not every vendor offered systems as they got paid off by intel. I remember helping a friend with building a kernel for one of the first available Opteron setups - that thing was impressive.
And then there’s the whole 64bit thing which intel eventually had to license from AMD.
Most of the big CPU innovations (at least in x86 space) of the last decade were by AMD - and the chiplet design of ryzen is just another one.
That’s already the friendly variant. Traditional find has a mandatory path as first argument, so to find in the current directory you need to do find .
It also doesn’t know if it really is a path - it just prints that as a likely error. You might just have messed up quoting an argument.
One fascinating example is one owner that replaced the DC barrel jack with a USB-C port, so they could utilize USB-PD for external power.
Oddly enough that’s also an example for bad design in that notebook: The barrel jack is soldered in. With a module that is plugged into the board that’d be significantly easier to replace - and also provide strain relief for power jack abuse. All my old thinkpads were trivial to move to USB-C PD because they use a separate power jack with attached cable.
The transparent bottom also isn’t very functional - it is pretty annoying to remove and put back, due to the large amount of screws required. For a notebook designed for tinkering I’d have wanted some kind of quick release for that. Also annoying is the lack of USB ports on the board - there’s enough space to integrate a USB hub, but just doing that on the board and providing extra ports would’ve been way more sensible.
The CPU module also is a bit of a mixed bag - it pretty much is designed for the first module they developed, and later modules don’t have full support for the existing ports. I was expecting that, though - many projects trying to offer that kind of modular upgrade path run into that sooner or later, and for that kind of small project with all its teething problems ‘sooner’ was to be expected. It still is very interesting for some prototyping needs - but that’s mostly companies or very dedicated hackers, not the average linux user.
Admittedly I’m just toying around for entertainment purposes - but I didn’t really have any problems of getting anything I wanted to try out with rocm support. Bigger annoyance was different projects targetting specific distributions or specific software versions (mostly ancient python), but as I’m doing everything in containers anyway that also was manageable.
For AI and compute… They’re far behind. CUDA just wins. I hope a joint standard will be coming up soon, but until then Nvidia wins
I got a W6800 recently. I know a nvidia model of the same generation would be faster for AI - but that thing is fast enough to run stable diffusion variants with high resolution pictures locally without getting too annoyed.
No, most companies also have mostly incompetent engineers.
Roku always was a company with great engineers and shitty money grabbing management. The new user creation always requested data not necessary for basic operation.
I find this situation rather entertaining. It shows yet again how important it is to educate people on the basics of how LLM work, including how they are being executed - I’m guessing with just a tiny bit more knowledge it’d also have been obvious nonsense to you.
Did pretty much the same with a new server recently - spent ages debugging why it didn’t find the SAS disks. Turns out, disks like to have power connected, and no amount of debugging on software level will help you.