I was curious what I had so I went through my box of old GPUs. It was a fun trip down memory lane.
I was surprised I didn’t find anything older. I must have gotten rid of them in the past. I know I had I had several Matrox Millennium cards at one time. I have an old DOS gaming machine that I haven’t touched in a while that probably has an older card in it. I am pretty sure there is a 3dfx card in it but I was too lazy to pull it down off the shelf though.
The RX 6700 XT as well as the RX Vega 64 still have some live in them and there is definitely some kiddo on a tight budget out there that would love to upgrade to such an GPU
In a certain aspect, it’s ridiculous that an almost 10 years old GPU such as the GTX 1080 are sold at an price higher than $50.
Gods.. I still remember by first real 3d graphics card, before the term GPU was even a thing, - seeing Quake/Quake2 in hardware acceleration was absolutely jaw-dropping, and then came the original Unreal..
I went from :
S3Virge (this was just 2d if I remember) →
3dfx Voodoo 1 →
3dfx Voodoo 2 →
Rage something? →
Riva TNT - I think this was the one that blew mind on Unreal, might have been the TNT2 →
Geforce2 →
Radeon 8000
The rest I can’t remember, and there was definitely a gap as I travelled around places, but there were so many great experiences back then that looking back, you’d almost laugh at them, except they were magical. Staying home all through Christmas to play X-Wing (the original version).. good times!
I still happily run a Vega 56, released in 2017. Wouldn’t consider it outdated at all. Well, except for when I want to work with Pytorch for uni, my card isn’t supported anymore But some ML applications still run?
I can even run a small LLM (7bn, quantized) at a usable pace. God I love the HBM memory. The little card really punches above its weight.
Have no fears, the 6700 XT will be getting used soon.
I wanted to build a test machine to do some testing. Which is what one does with a test machine I suppose…
However, some of my testing is related to gaming and so I needed something newer than any of the MBs I had lying around. I went to buy an entry-level setup and then remembered about the current memory-pocalypse. I realized that one of my older boards had DDR4 in it so I picked up a Ryzen 5600x and a b550 board to go with it. I already have everything else I needed for the build available. Well, as long as I can find the AM4 mounting hardware for my cooler.
I’m on an AM4 Ryzen 5700x (8 core); but want a 5950X 16 core to replace it (anyone selling one for a low price ), but still using an MSI NVIDIA GEforce GTX1050 OC (2Gb GDDR 5) in it:
So glad I’m not a gamer and do mostly CPU intense computing, so I can’t see any gain from anything faster and more powerful for now, and probably wouldn’t even notice any difference. I am although reading up on and considering running a minor local and task specific AI, and a new GPU may just come in handy soon, but at the current rate of DDR prices which will put the GPU’s, less plentiful, and at a much higher price, I think I will hold back and drag it out until my GPU or I die if the big tech AI insanity doesn’t stop!
Just a few month ago you could easily get a AMD Instinct MI50 GPU with 32GB (or 16GB) of HBM2 VRAM at reasonable prices. Might not be the fastest in terms of token generation, as it is based on the Vega architecture. But for a price of $225 for 32GBs of VRAM not too bad. Especially due to it’s 4096-bit memory bus width that achieves up to 1024 GB/s bandwidth. The only downside would be that it is designed for server uses and doesn’t have an active cooling solution.
Only the recent developments drove the price up to $325 or so. Which is pretty close to the current price for a RX 9060 XT with 16GB of VRAM. But that`s a much more recent card which only has a 128-bit memory bus and only 1/3 the memory bandwidth that would top out at 322.3 GB/s memory bandwidth.
In terms of larger AI models that would require more than 16 GBs of VRAM, You’ll barely find a new GPU with 32GBs for less than $1000 currently. And I guess most of the MI50s already found their way to those which are into self-hosting.
So you are speaking of a used server native card, right? I am unfamiliar with the AMD Instinct MI50 GPU with 32GB (or 16GB) of HBM2 VRAM, and server centric GPU’s in general other than the differing tech and requirements for AI, some you mentioned.
Sadly if I had even $250 to spend in the last few years I would have got a new card quite a while ago when I built the system, or within a few months after, since I already went over budget due to finding a better motherboard on sale for just a little more than the one I picked, one or two Items I had picked being out of stock (I got it all at Micro Center) … all for a better system at least, and why I just used the old card for the time being which due to the insane price increases including for RAM that has been sitting in warehouses and on the shelf for a few years already I see as a terrible time to buy. I’m glad that I have plenty of real I to hold off on AI indefinitely which is still way more A than I anyhow, as my recent experiments with them has only shown far worse than I thought, so whoever is claiming it’s great, isn’t using it to find facts, not giving it precise instructions, nor a clue what answers to expect, not looking for true answers but confirmation bias… IOW not very intelligent themselves, or monetarily invested in it and trying to increase their profits…
With Windows 10 end of life, and 11’s hardware requirements and compatibility issues, more used computers perfectly well suited for Linux are already for sale used with many more to come, many with RAM and GUI included, for their owner not knowing tech stuff, and not up on the market or what things are currently selling for, I got lucky that way for every computer I bought used, it’s not as much about what you buy used, as it is who to buy from like more well to do unlikely over-clockers or heavy users…
I have two PCIe 4.0 x 16 slots, so two lower end cards at a low price can add up to more performance than single better and newer card that costed more to begin with and selling for more than the two with, the prices still jacked after the GPU/bitcoin mining debacle and never went down anywhere near as much as they should have due to corporate greed, which here in the US is now standard procedure to cash in big time and double dip.
I wish I had kept every medium I ever used from A: drives [what a great sound] to Iomega Zip [what a fantastic sound that one made too. 2.5” before my times and hard drives with 100MBs…do you have a box of that stuff ?
This has gone somewhat OT but I keep looking at USB floppy drives for my Amiga emulation PC. It’s not even the drive noise I miss (this can be emulated in WinUAE). I think it’s the click of the mechanism taking the disk. They are all slimline though which puts me off and even disks that have been stored well might be past it from what I’ve read.
Back on topic : I’ve been looking at AGP cards for my new PentiumIII system and the decent ones go for a fair few hundred quid tbh.
I was just browsing tbh. I haven’t even tested the PC yet, though it is sold as working.
I initially bought it with plans on running a few DOS games that are too CPU intensive for DosBox on my PC. Had the thought of running Windows but not sure if I would run Win9x or XP on it which will be a big factor in what GPU to use.
I will be very limited as the case will only take a low profile, single slot card and the PSU might not be able to handle very much. (not swap-able as it’s a Dell).
It was going cheap enough that if it doesn’t work I probably won’t bother returning it but I need to do a bunch of research before upgrading it.
But some ML applications still run?
