NVIDIA GeForce RTX 50-series Blackwell Availability

Today, just like the crypto-mining mania, there's this "AI" mania. I know people who are buying 5090 not because they are gamers but because they want to do edge AI. They are also more likely to be willing to pay more. I also heard some rumors about some people in China are replacing 5090's memory with chips of twice the size, making them 64GB, and also some rumors of the 3 bits modules with potential to 96GB. Together with the ban of AI chips to China, I guess this is just going to get even worse.
There are 4GB GDDR7 chips?
 
Technically the standard goes up to 8GB/chip but right now we're getting only a small number of expensive 3GB chips.
4, 6 and 8 should all be possible in the future though.
That's what I thought. I'm wondering how are they modding 64GB on a 5090? I thought there had to be some PCB considerations to do clamshell.
 
There are 4GB GDDR7 chips?

I'm guessing it's referring to doubling VRAM by making them double sided, not using 4GB chips which don't exist.

While the standard allows for it whether or not it's technically and/or financially practical for them to be deployed in GDDR7's lifespan would be debatable. GDDR6 in spec allowed for 1GB, 1.5GB, 2GB, 3GB, and 4GB. Only 1GB and 2GB were ever deployed in market.

That's what I thought. I'm wondering how are they modding 64GB on a 5090? I thought there had to be some PCB considerations to do clamshell.

I think it should be kept in mind with China we aren't talking about just modding these at home with just a soldering iron or something. It's companies with proper equipment and trained professionals. They've already done it at scale with the 4xxx series.
 
I'm guessing it's referring to doubling VRAM by making them double sided, not using 4GB chips which don't exist.

While the standard allows for it whether or not it's technically and/or financially practical for them to be deployed in GDDR7's lifespan would be debatable. GDDR6 in spec allowed for 1GB, 1.5GB, 2GB, 3GB, and 4GB. Only 1GB and 2GB were ever deployed in market.



I think it should be kept in mind with China we aren't talking about just modding these at home with just a soldering iron or something. It's companies with proper equipment and trained professionals. They've already done it at scale with the 4xxx series.
They are making custom PCBs? And wouldn't the firmware also have to be changed? Or does it autodetect when 2 modules are present on a channel?
 
They are making custom PCBs? And wouldn't the firmware also have to be changed? Or does it autodetect when 2 modules are present on a channel?

I haven't looked into this in detail. From what I understand is that this became publicly aware when companies started making them available to the public. It's not a situation in which some enthusiast was modding them and documented how they did it exactly.



In terms of the firmware it's worth noting that Nvidia's firmware signing I believe was broken (at least some degree) in 2023. So firmware level modifications being a requirement may not be a barrier in itself.

 
Back
Top