-

Status
Not open for further replies.
xenium said:
'Preciate it man. Surprised it hasn't sold yet at this price. Might be a sign to keep the thing afterall..
Click to expand...
Yeah. I keep telling myself I have too many video cards, but the price is seriously tempting me.
 
How many 4k plex streams can you run on a A4000 over the P2000? Can you slave a A4000 together with a 3090 to get 40gb of ram for AI to play with or does that have to be like cards?
 
Kirika said:
Can you slave a A4000 together with a 3090 to get 40gb of ram for AI to play with or does that have to be like cards?
Click to expand...
you can but the slowest card will bottle neck the faster one.
mixing a Quadro with a 3090 might not play nice because of drivers
https://www.reddit.com/r/LocalLLaMA/
You can always build 2 machines
Quadro for inference ( server) and 3090 for training/gaming/ work/personal stuff
Nobody buys Quafro for training ( memory is too slow) but they are good for inference
 
Kirika said:
How many 4k plex streams can you run on a A4000 over the P2000? Can you slave a A4000 together with a 3090 to get 40gb of ram for AI to play with or does that have to be like cards?
Click to expand...
It's going to depend on the video(s) in question, I'd guess 10-15 transcodes? It would be a beast for that task. Also you can run just about any two supported Nvidia GPUs together for some ML tasks - but a headache when/if different libraries needed for different architectures. Since A4000 and 3090 are both Ampere, it should be relatively painless. As Binar said, not great for training but since this one has been beefed up (effectively a 3070/ti) it would still be better than a normal A4000. For just running local models, the biggest "problem" with that combo is that 40gb VRAM total. The best models are the 70b+ ones, and they require just a bit more than 40gb VRAM to run (mid 40s for roughly realtime feedback) - so you'd be able to do a lot more, but still be mostly limited to the roughly 30b model size. That said, still a LOT of cool ML stuff you can do with two powerful GPUs in tandem.
 
I keep looking, but just doesn't make sense because I actually have a server chassis so I'd have to put half of it back, which defeats the purpose of a very nicely modified card. Free bump though, great work.
 
Ready4Dis said:
I keep looking, but just doesn't make sense because I actually have a server chassis so I'd have to put half of it back, which defeats the purpose of a very nicely modified card. Free bump though, great work.
Click to expand...
Not sure I understand…

I mean, if you buy this, then I can be sure I won’t. There’s nothing to unmod here. The original HSF won’t work with an 8 pin power connector. The 8 pin is necessary for higher max power (and, thus, higher performance). Not sure I see a reason you’d have to put anything back. What am I missing?
 
PurduEE said:
Not sure I understand…

I mean, if you buy this, then I can be sure I won’t. There’s nothing to unmod here. The original HSF won’t work with an 8 pin power connector. The 8 pin is necessary for higher max power (and, thus, higher performance). Not sure I see a reason you’d have to put anything back. What am I missing?
Click to expand...
I'm assuming he only has 1-2 free slots in said server chassis? I've actually had 2 other people inquire about it in the exact same position. The card is 2.25 slots, and really needs 3 full slots to let the HSF breathe. Which is probably 99% of typical ATX/mATX builds, and a lot of bigger ITX cases too. Exactly as the card was built for (my dual-purpose ITX rig, for gaming/workstation use).

But yes, it would be a lot of work to de-mod the card (not to mention the stock HSF is long gone). I had no idea there would be so much demand for the crappy single slot cooler configuration.
 
PurduEE said:
Not sure I understand…

I mean, if you buy this, then I can be sure I won’t. There’s nothing to unmod here. The original HSF won’t work with an 8 pin power connector. The 8 pin is necessary for higher max power (and, thus, higher performance). Not sure I see a reason you’d have to put anything back. What am I missing?
Click to expand...
As xenium pointed out, the size of the card in a server chassis wouldn't fit well, so I would have to switch back to the orignal heat sink and remove the fan... (or 3d print a new cover with a blower fan) which defeats the entire purpose. It's a sweet card, just won't work for my use case (plus i see the fan header cable, not sure if that plugs into a MB fan header port, which the server doesn't have, so no fan control either without hacking my server).
 
Ready4Dis said:
As xenium pointed out, the size of the card in a server chassis wouldn't fit well, so I would have to switch back to the orignal heat sink and remove the fan... (or 3d print a new cover with a blower fan) which defeats the entire purpose. It's a sweet card, just won't work for my use case (plus i see the fan header cable, not sure if that plugs into a MB fan header port, which the server doesn't have, so no fan control either without hacking my server).
Click to expand...
It's just a typical 4 pin pwm fan connector. Like I said in the OP as well, if you buy a cheap crj > 4 pin pwm adapter off Amazon the card will control the fan no problem. It just has an annoying fan curve, stock.

The fan is powered/controlled externally. I plugged it into a spare mobo fan header, and used the FanControl program in windows to set a custom fan curve.

If desired, you can use a cheap crj > 4 pin connector and let the card itself control the fan. But the stock fan curve is very aggressive (for the stock blower fan), and has a minimum fan speed of 40% that can't be overridden.
Click to expand...
 
gimme-what-about-bob.gif
 
Status
Not open for further replies.
Back
Top