NVidia GeForce RTX 50x0 cards - 70% performance increase, but AI > you
- Thread starter Izo
- Start date
RobXIII
Urinal Cake Consumption King
- 4,573
- 3,901
Which Pimax do you have, and how do you like it?
I almost pre-ordered the Pimax Dream Air but questions on when it would ship, and the weird 'charge you extra for a service ..thing' on their website wasn't clear so I backed out for now.
Noodleface
A Mod Real Quick
- 39,758
- 18,447
Hekotat
FoH nuclear response team
- 13,308
- 13,781
Sim Racing. The only permanent use case I've found for it (and yes it's incredible).
Leadsalad
Cis-XYite-Nationalist
- 6,608
- 14,972
Crystal Super OLED. It's okay, it's the only headset I've had long term experience with. The lenses definitely have some off central point chromatic aberration and you're still definitely looking though goggles.
Really my complaint right now is just the amount of processing power needed to keep it at 90hz. I can't get Assetto Corsa to run well at all compared to LMU.
Haus
I am Big Balls!
- 20,573
- 82,964
So today I find myself at a quandry....
Decided to do the "tiny" upgrade to the desktops in the house. (my rig and the one
Mrs. Haus
uses). 3060s with 8GB to 5060TIs with 16G. For the first time I can remember I'm doing an upgrade NOT because of gaming performance. But rather that we're both leveraging AI stuff and need the memory on the cards to be high enough to run some more capable models.
Both our rigs are running 750w "bronze" rated power supplies. And my calculations show I could PROBABLY get away with running both cards in the machines (both our existing, and the 5060 along side) but it would be borderline.
Now I'm debating that, versus taking both old cards back to the lab to drop in my lab box and have it run dual and have those running something....
Decided to do the "tiny" upgrade to the desktops in the house. (my rig and the one
Mrs. Haus
uses). 3060s with 8GB to 5060TIs with 16G. For the first time I can remember I'm doing an upgrade NOT because of gaming performance. But rather that we're both leveraging AI stuff and need the memory on the cards to be high enough to run some more capable models. Both our rigs are running 750w "bronze" rated power supplies. And my calculations show I could PROBABLY get away with running both cards in the machines (both our existing, and the 5060 along side) but it would be borderline.
Now I'm debating that, versus taking both old cards back to the lab to drop in my lab box and have it run dual and have those running something....
Hateyou
Not Great, Not Terrible
- 19,000
- 51,442
Hasn’t running two cards been pointless/worse than running one for like 20 years now?So today I find myself at a quandry....
Decided to do the "tiny" upgrade to the desktops in the house. (my rig and the one
Both our rigs are running 750w "bronze" rated power supplies. And my calculations show I could PROBABLY get away with running both cards in the machines (both our existing, and the 5060 along side) but it would be borderline.
Now I'm debating that, versus taking both old cards back to the lab to drop in my lab box and have it run dual and have those running something....
- 1
Noodleface
A Mod Real Quick
- 39,758
- 18,447
Haus
I am Big Balls!
- 20,573
- 82,964
The way I see it, the games my wife and I like playing all play reasonably on the cards we already have. If the 5060s are just used for the AI work that perfectly serves my purposes well. For now, I think I'll remove the old cards, upgrade in place. Then once I can hit the store or find a good sale on PSU upgrades, I'll upgrade them to 1000W PSUs and drop the old cards back in.
Then monitors get plugged into the old cards, new cards just run AI workloads and lean on the old cards as needed.
Jovec
?
- 947
- 562
Not sure I follow. You want to run thee 5060ti 16GB for AI while having the older 3060 in the same box for gaming? 750w should be enough, even bronze rated, assuming a decent make and model. 3060 and 5060 should both be around 200-225w each under full load, and most CPUs when gaming are under 100w. Add another 50W for a single drive, fans, rgb, etc, and you should be under 650W, if not under 600W. Unless your CPU is a 14900K or similar...
Another option is to keep your existing systems, then drop both 5060ti 16GBs into another system that you both can share remotely for AI and have access to 32GB of VRAM for your models.
Another option is to keep your existing systems, then drop both 5060ti 16GBs into another system that you both can share remotely for AI and have access to 32GB of VRAM for your models.
Chapell
Trakanon Raider
- 651
- 760
Control
Golden Baronet of the Realm
- 5,842
- 16,280
could also look for used 3090sSo today I find myself at a quandry....
Decided to do the "tiny" upgrade to the desktops in the house. (my rig and the one
Both our rigs are running 750w "bronze" rated power supplies. And my calculations show I could PROBABLY get away with running both cards in the machines (both our existing, and the 5060 along side) but it would be borderline.
Now I'm debating that, versus taking both old cards back to the lab to drop in my lab box and have it run dual and have those running something....
Hateyou
Not Great, Not Terrible
- 19,000
- 51,442
Fucker
Log Wizard
- 16,600
- 41,129
Not only that, but at full power. Likely replaced because they started popping errors all over the place.
Electromigration - Wikipedia
Control
Golden Baronet of the Realm
- 5,842
- 16,280
This is true. Still the best value for AI usage though (possibly outside of setups that can leverage old server cards).
Fucker
Log Wizard
- 16,600
- 41,129
IMO, you are better off with a paid service. I've done quite a bit of local AI stuff, and mostly have moved off of it. This is on Apple, AMD AI, and Nvidia stuff.So today I find myself at a quandry....
Decided to do the "tiny" upgrade to the desktops in the house. (my rig and the one
Both our rigs are running 750w "bronze" rated power supplies. And my calculations show I could PROBABLY get away with running both cards in the machines (both our existing, and the 5060 along side) but it would be borderline.
Now I'm debating that, versus taking both old cards back to the lab to drop in my lab box and have it run dual and have those running something....
5060TI has the memory, but neither the processing power nor the memory bandwidth to do anything useful with them. Same with Apple "M" chips, same with AMD Ryzen "AI" garbage.
Not only that, but tools for AI constantly break. AMD ROCm is a sad joke, and managing packages and installs on even Nvidia stuff can be a tedious affair. I wasn't paying attention, and updated the package on one of my apps, and it took the WHOLE thing down with it. None of my other apps worked, and had to be reinstalled from scratch.
Imagine doing this with Apple stuff, AMD stuff, and Nvidia stuff.
Now, I just use online apps.
If you want to futz with it, it is fine. That's what I started out doing. In that train of thought, you'd be better off with a 5070ti. Double the memory bandwidth of a 5060ti, Apple "M" or Ryzen "AI". Same-ish cost of 2 5060ti's, but WAY more performance.
Haus
I am Big Balls!
- 20,573
- 82,964
Well, in this case it was 2 so that we'd both basically get our "once every 5 years or so" gaming rig upgrade in while also doing AI stuff.
So far on hers I just put AnythingLLM on it and have it running Gemma4-26B (pulled via a GGUF re-quant) and it's running well for what she wants to do. (i.e. she ran it and said "Yeah, that works!" so I'm happy). Her only annoyance was in the makefile for it I told it to refer to her as Princess Gassypants.
If figured with these smaller upgrades it let's me start tinkering. If I'm going to get more "super serious" I'm looking at ways I can get my work to partially subsidize me something alone the lines of a Nvidia Spark box. Or at minimum step up to one of those 32G sporting 5090s, and build it out in a dedicated box in the lab. But that's down the road if this pans out enough for now. I just know some guys who when they started doing the type of "running on things all the times and agentic enabled shit" game they pulled back a three digit monthly bill from Claude and that's a recurring cost I'd prefer to avoid.