The analysis comparing GPU flagships and price/wattage is somewhat shallow because it hides another thing that has been going on, the reduction of fundamental specs of the various classes of card. It used to be the case that a x80 chip meant it had a 512 bit memory bus, that it was ~500mm^2 die and hence a fully maxed out GPU and you got all that for ~$400-500 or so. Then the 680 came out and its specs were more like an x70 card from the prior generations, its only 294mm^2 and a 256bit bus, it was a rename of the lower class card, they never produced a flagship for that particular generation and some of that degradation in specs carried over to the 780 ti as well which now only had a 384 bit bus but its die size was ~561mm^2.

An RTX580 is now 378mm^2, which is about an x70 in pre 2012 terms and 256bit bus which is also x60-x70 class.

That process has resulted in the titans and the x90 and the x80 ti all slipping above the x80 as its specs gradually decline and its price is still going up compared to the historical picture, enormously more than inflation. During the same period CPUs on the other hand have stayed fairly similar in price with a steady increase in performance at a price point. That 680 oddity in the historical area was the moment things changed and AMD had a big part to play in the reason why with their 7970 being priced so much higher. This process started then in 2012 and its been getting worse as time goes on.

I keep trying to use the various LLMs that people recommend for coding for various tasks and it doesn’t just get things wrong. I have been doing quite a bit of embedded work recently and some of the designs it comes up with would cause electrical fires, its that bad. Where the earlier versions would be like “oh yes that is wrong let me correct it…” then often get it wrong again the new ones will confidently tell you that you are wrong. When you tell them it set on fire they just don’t change.

I don’t get it I feel like all these people claiming success with them are just not very discerning about the quality of the code it produces or worse just don’t know any better.

Would be a nightmare to adminster as well, has so much less automation and tooling for deployment and updating of software. Even now the updating of apps on windows is a mess and the closest they have come is winget that centralises the entire thing through stores, completely useless for the corporate world. There is a reason Linux won on the server.

Pretty certain cd and pwd have changed over the years. The kernel hasn’t remained the same so the commands that use it wont and now we have faster methods to do various things like getting file data the commands that depend on it will change. Less quickly than something that is still gaining features but bit rot is a very real effect since every single part of software is in constant flux.

The fact it recommends popular stuff is a useful addon feature, its a good way to look at what others are watching.

Its a big problem. I also dump projects that don’t automatically migrate their own SQLite scehema’s requiring manual intervention. That is a terrible way to treat the customer, just update the file. Separate databases always run into versioning issues at some point and require manual intervention and data migration and its a massive waste of the users time.

31 Containers in all. I have been up as high as ~60 and have paired it back removing the things I wasn’t using.

I also tend to remove anything that uses appreciable CPU at idle and I rarely run applications that require further containers in a stack just to boot, my needs aren’t that heavy.

I reject a lot of apps that require a docker compose that contains a database and caching infrastructure etc. All I need is the process and they ought to use SQLite by default because my needs are not going to exceed its capabilities. A lot of these self hosted apps are being overbuilt and coming without defaults or poor defaults and causing a lot of extra work to deploy them.

The only detail really is that at least 2 of the N machines you are using have to be on at the time so where ever a change was made is synced to another machine that is on and this continues so that you never end up booting a machine to use when nothing else with the latest files is available. This is where having a centralised low power machine is valuable and saves having a desktop or a laptop on when it doesn’t need to be.

I really wish the desktop version of the world had not become so marginalised as local programs are considerably better to use than websites, they are so much quicker, accessible and easier to use.

On the one hand they were talking selfhosting and then they pull out multiple $10s thousands rack servers. People don’t need a data centre at home to sync some files, pictures, email and play some media!

Every one always says XMPP and there were a lot of recommendations for ejabberd. I tried this recently and it was a total disaster, I do not have a working chat server. If I followed the docker instructions the server would just crash with no details of what went wrong. Where it should have been creating a default server config file it was instead creating a directory with the wrong permissions then promptly crashing. I tried following their documentation but after about 6 hours of messing about and adding more and more I still couldn’t get a client to login to it. I have no idea how to make this work.

So whatever the solution ultimately is I can’t recommend Ejabberd.

Initially a lot of the AI was getting trained on lower class GPUs and none of these AI special cards/blades existed. The problem is that the problems are quite large and hence require a lot of VRAM to work on or you split it and pay enormous latency penalties going across the network. Putting it all into one giant package costs a lot more but it also performs a lot better, because AI is not an embarrassingly parallel problem that can be easily split across many GPUs without penalty. So the goal is often to reduce the number of GPUs you need to get a result quickly enough and it brings its own set of problems of power density in server racks.

I use a 5600g on b450 ITX board and 4x 8GB Seagate drives and see about 35W idle and about 40W average. It used to be 45W because I was forced to use a GPU in addition to a 3600 to boot (even though its headless, just a bad bios setup that I can’t fix) and getting a CPU with graphics dropped my idle consumption quite a bit. I suspect the extra wattage for your machine is probably the bigger motherboard and the less efficient CPU.

It is possible to get the machine part down into single digits wattage and then about 5W a drive is the floor without spinning them down, so the minimum you could likely see with a much less powerful CPU is about 30-35W.

Make sure none of the exceptions are ticked and the Minimum number of articles to keep per feed is also 25 or below. Then its up to the cron when that runs so you might have to manually purge it and optimise the database to see what it will actually keep.

I can’t say I have ever worried about it, been running FreshRSS for years and it seems to keep its database size in check fairly well and the defaults have worked fine for me and it rarely gets above 100MB. So I know it “loosely” works in that old articles are absolutely getting purged in time but have no idea how strictly it follows these rules.

The DMZ for the ISPs router forward to the second router, then everything that hits your outside IP will be forwarded to router 2. Then on Router 2 you open the ports for your service and forward to the internal machine. That should all work fine.

Its quite complicated to setup as well, just went through the instructions and its a long way from just add to docker and run unfortunately. Would be nice to be able to just get a runner in the same or different docker and it just works easily without a lot of manual setup in Linux of directories and users and pipes etc.

I did the same move from contabo to Netcup. Contabo I had all sorts of weird bandwidth limiting problems that I couldn’t explain and which the continued to deny they were throttling. Netcup worked perfectly.

The problem is the information asymmetry, there is always another person for a fraudulent company to exploit due to a dysfunctionally expensive court system. Its why we need market level regulations and public institutions that recover peoples money and fine the organisations for their breaches. This sort of thing works a lot better in the EU than in the US due to the sales laws, the ability to return within 2 weeks, default warranty on goods out to 12 months and expectations of goods to be as advertised forced onto the retailers. They work, they need more enforcement from regulatory bodies but retailers do follow them for the most part and quickly change tune when you go to take legal action when they don’t because courts know these laws inside and out.