Non Forced 2019 Upgrade Edition AMD 3950x

[Aslanna]

A few recent upgrades lately but nothing too fancy. All of them have been storage related.

Replaced an 8 TB drive with an 18TB. That is basically my Plex drive so I have some room to grow now as I was down to a few hundred GB free.

Replaced a 250GB SSD (bought in 2014 for $110) with a 1TB SSD ($53 in 2023). This is my D drive where I install apps. Need to keep C clean!

Added a 4TB SSD to copy half off a 4TB WD Blue drive. Will make that an AI drive with the Blue being dedicated to emulation stuff.

I figured with SSD prices only forecast to increase now would be a decent time to buy. I should probably replace the Blue with another 4TB SSD which would then leave me with only one mechanical drive. Maybe if they go on sale a bit. It was $250 for a Samsung 870 EVO 4TB which I don’t think was too bad.

As far as any other upgrades I am skipping Intel 14th gen which I was planning to go to prior to them announcing it so I’m hoping 15th gen will be worth it.

[Winnow]

I figured with SSD prices only forecast to increase now would be a decent time to buy.

I thought about getting an 8TB SSD but still seems a little pricey. I still have 38TB of spinning drives I'd like to replace one old 8TB HS in particular seems like it's on its last legs. I've got 9 TB of SSD drives so far. Hate spinning drives. I'm out of room. I'd like to transfer to SSDs but whatever I'd buy would probably be used for expansion. One thing I need is a bigger AI drive. I have all my AI stuff Stable diffusion/LLM and all the other python apps on a 2TB SSD and it's not enough. Needs to be at least 4TB.

The external Samsung T7 Shield is a good 4TB drive ($249). USB C drives perform real well. All those bazillion USB ports I got on my last MB (pic earlier in this thread) are used! I might have one one USB port on the back when you add in the mice, keyboard, gamepad, etc etc and the MB has 12 USB ports total.

[Aslanna]

My 4TB magnetic was split between ai and emulation so I designated the new SSD for ai with the magnetic only having emulation stuff. At some point I may replace that with an ssd. I have been hearing some issues with the 870 EVO though so that is a bit concerning. Hopefully those have been resolved with firmware updates.

My next big storage update will probably be setting up a NAS. I’ve had that on the list for awhile but have kept putting it off.

[Winnow]

I have been hearing some issues with the 870 EVO though so that is a bit concerning. Hopefully those have been resolved with firmware updates.

All of my SSDs are Samsung. I know one is an 860 EVO (my only internal SSD other than my NVME system drive), not sure if 870 in particular has issues but all my Samsung SSDs have worked great. My most recent is the T7 Shield 4TB. Kind of like the rubbery case. All my external SSDs hang off short usb cables on the back of my case.

If I start upgrading my SSDs I'll have to replace the smallest ones which are 2TB already. I had to get the 4TB due to AI. LLM and Stable models/loras etc etc take up a lot of space. I just looked and I have 20 other "AI" related apps on my 4TB. I also have various python based Youtube/Video leaching apps. I have unlimited bandwidth so will sometimes download entire VR Youtube channels to watch on my Quest. It's more convenient to have them locally than stream them.

Offtopic:

Note: if you ever need to download youtube videos (highest resolution quality or otherwise) use this app:

https://github.com/ErrorFlynn/ytdlp-interface

No install necessary, just place the files from this zip into their own folder

https://github.com/ErrorFlynn/ytdlp-int ... terface.7z

Open it with ytdlp-interface.exe

You can update the latest yt-dlp and ffmpeg from within the app and also the latest version of the app itself. yt-dlp does the work, the interface is the UI that replaces all the tedious line commands. yt-dlp is pretty amazing. It actually supports hundreds of video websites for downloading but this interface focuses on youtube.

Super easy to use. You can download single videos or entire channels/playlists etc. Conveniently places videos from playlists and channels into their own folders. Youtube is starting to restrict high bps/fps 1080p/2160P unless premium paying customers. This app bypasses that. You can set your "preferred resolution to 2160 and it will grab the highest resolution available.

Any youtube ads are removed by default but there are some settings for sponsor blocking etc. I haven't had to mess with hat at all but they're there just in case.

You may not use it a lot but could be very useful for those oddball series etc you might want to save. You just paste the video link or playlist link or entire channel link and the app does the rest. I have a couple nice video players on my Quest 3 I can stream from my PC with much better controls than Youtube offers so definitely helps me to have the files local.

I use

Code: Select all

 %(upload_date>%Y-%m-%d)s - %(title)s.%(ext)s

as my Output template so that the file names start with the date of the video which makes for much better sorting and watching in sequence based on the file titles.

Also if you have a large channel and you want to download the oldest videos first, enter:

Code: Select all

--playlist-reverse

In "custom arguments" and it will start downloading oldest video first instead.

YT-DLP Interface is just one of many UIs for yt-dlp but by far the best IMO. On rare occasion I'll use yt-dlp-gui for a single large video because is uses Aria2 which splits of the video into hundreds of parts and downloads them all at once pretty much maxing out your bandwidth to download the video as fast as possible. Unless you're in a rush for a particular video, not normally needed but nice.

[Aslanna]

Search for 870 EVO and you'll see all sorts of issues with early failures. Also, and not 870 in particular, older files are super slow when it comes to transfer rates. I don't know if that has been resolved with firmware updates but there are things you can do to update your file dates so the system thinks they are 'new'.

[Winnow]

6 years 4 months

This motherboard/CPU combo has been fantastic.

I've upgraded the RAM multiple times and GPU multiple times but the motherboard and CPU have been workhorses. (and a great case)

Considering the state of home PC market, I may try for 10 years with this if it holds up.

I had a scare early this week 2 of my 4TB SSDs stopped working. $679 each at current prices. I paid $281 and $239 for them.

I rebooted and working fine for past 3 days so hopefully some power surge or something. I really don't want to pay for memory/storage anytime soon.

I did wise up and backed up my Everything Index db to another drive. It doesn't help much if the drive it's on crashes. It was a huge help when I lost a 14TB HD to see what I had on the drive so I could replace some things.

RTX PRO 6000 Max-Q is down to $9,399 so hopefully they will settle. I'm still making daily use of that extra VRAM. LTX 2.3 has turned out to be a great video model and a new video model was released a few days ago to try out. audio integration makes such an big impact.

[Aslanna]

You were warned that SSDs (and HDDs) would be going up so you should have bought some just in case. I picked up a few sizes including a 4 TB which I think was $270 at the time and is now at a MSRP of $655. Which is dumb. I picked up a 20 TB HDD for $350 which is about $500 now although those didn't rise too much so far.

As far as storage goes I'm set for many years. Hopefully by the time I actually need to buy new stuff it will be reasonably priced.

[Winnow]

You were warned that SSDs (and HDDs) would be going up so you should have bought some just in case. I picked up a few sizes including a 4 TB which I think was $270 at the time and is now at a MSRP of $655. Which is dumb. I picked up a 20 TB HDD for $350 which is about $500 now although those didn't rise too much so far.

As far as storage goes I'm set for many years. Hopefully by the time I actually need to buy new stuff it will be reasonably priced.

I'll be OK on storage. I have a huge porn buffer (although a lot of that was lost with the 14 TB crashed HD awhile back. I can always delete some of that to make room. Also lots of TV series/movies stored. I don't really watch/download porn anymore. I can make my own easy enough with AI. Not sure what the AI haters are smoking but for porn at least, AI is better...and isn't that always the case? New tech = porn first then the mainstream follows.

I've also got assloads of ASMR videos I could delete. I download entire channels but after listening to them there's only a handful of videos that I listen to repeatedly that put me to sleep.

AI images and videos don't take much room. (the models and loras take a lot of room though). I've got 10's of thousands of videos but they are like 2-4mb each. No idea on AI images, maybe half a million of those or something. I don't batch process images. I create them one by one almost always. Sometimes I'll set up a prompt "storyboard" with like 6-8 prompts and run that to create an entire scenario. Or for first/middle/last frames for videos.

Key thing about the AI images and videos is using Diffusion Toolkit. All those are thumb-nailed and indexed by prompt keywords, model, loras etc. I have my most recent 100K images and videos loaded into it. Easy to reference old images and videos for prompts, image to image image to video etc. You can also just drag the image or video right into comfy to load up the workflow with prompts, loras, etc.

Text LLM models can be huge in size but as better ones come out, they get replaced. no need to hold onto older versions usually.

qwen3.5-27b-uncensored-hauhaucs-aggressive

is the best model out right now for all purpose. I use the Q6_K (23 GB) and like 132,000 context. It's also a vision model which makes it even more useful.

qwen3.5-9b-uncensored-hauhaucs-aggressive (10.4GB) is also very good and that's only 10GB Q_8 so easy to fit and keep a high context as well. You want to keep at least 128K context memory for Qwen3.5 models to perform well. The rest of the parameters are listed in the model notes. (temp, Top P, Min P, Top K Sampling etc). It's important to set those correctly for each model. 128K context is really good size so the model doesn't forget your conversation for a long time.

I can run 120B models quantized but the 27B is the sweet spot.

LM Studio is still the easiest way to run models. just cut and paste the above model in it's search and download and you're up and running. (after setting the parameters that are listed in sidebar the notes when you download models)