> They do this so you'll have to buy several cards for your AI workstation.

AFAIK you can't do that with newer consumer cards, which is why this became an annoyance. Even a RTX 4070 Ti with its 12 GB would be fine, if you could easily stack a bunch of them like you used to be able with older cards.

It's "easy" if you have a place to build an open frame rig with riser cables and whatnot. I can't do that, so I'm going the single slot waterblock route, which unfortunately rules out 3090s due to the memory on the back side of the PCB. It's very frustrating.

I think parents point is that NVLink no longer ships with consumer cards. Before you could buy two cards + a cable between them, and software can treat them as one card. Today you need software support for splitting between the cards, unless you go for "professional" cards or whatever they call them.

Maybe that's what they meant, and it'd be cool if nvidia still offered that on consumer cards, but thankfully you don't need it for LLM inference. The traffic between cards is very small.

Isn't the issue that the software needs to explicitly add support for it now, compared to yester-yesterday when you could just treat them as one in software?

There was a rumor that 5090 or 5090D for China may or may not come with multi-GPU software locked. I think GP's referring to that. It's not clear if it is the case with retail cards.

I honestly don’t know why people aren’t more upset by this and still get on their knees for Nvidia. They made the decision specifically to cripple consumer card memory because they didn’t like data centers were using them instead of buying their overpriced enterprise cards that were less performant. They removed NVLink because people were getting better performance out of their two $400 cards than the $1,500 cards Nvidia was trying to peddle. They willfully screw consumers and people love them for it.

Because sensible people just use the cloud at this point, you can probably get several years of training for $6000

It buys you approximately two days (with reservation discount) of a single p5.48xlarge instance, which has 2TB of RAM, and 640GB of VRAM in 8x H100 cards. In fact that is the pricing example they use: https://aws.amazon.com/ec2/capacityblocks/pricing/

MI300X (RunPod) 192gb ram Hourly Rate: $2.49/hr. Break-even Point: You can rent for 2,410 hours (~100 days of non-stop-continuous use) before reaching the cost of the $6000 Mac. Mac's top out at 192GB not 2TB ;) Consideration: If your AI training requires sporadic use (e.g., a few hours daily or weekly), renting is significantly cheaper. MI300X will also get you result many times faster too, so you could probably multiply that 100 days!

The 5090 is 32GB out of the box. Not that that's anywhere near the top of what you can do on an Apple, but at least it's movement.

Or buy 2 Nvidia digits for $6,000 to get 256GB vram.