NVIDIA GeForce RTX 3090 Ampere GPU to sport 4,992 CUDA cores and 12 GB of 18 Gbps VRAM

I just looked up some 8 Ghz GDDR6 on digikey for curiosity's sake. These appear to be 512 Mb ICs vs say the 1Gb on a 2080 Ti? so no idea how that would impact the price, but still pretty cheap compared to the numbers I've seen floating around above, and give you an idea about bulk pricing when the minimum quantity is 2000.
View attachment 234682

Unless I’m reading it wrong that is a heck of a lot more than the numbers I provided. That’s showing a per unit price of $24.50 and $24.80 for tape & reel and tray respectively, with a minimum order of 2,000 units. The numbers I found said $11.65 per unit for 1GB ICs.
 
Talking about the 1080ti being future proof. Here is another example of it being future proof in a brand new 2020 title. Look at how that 2060 super keeps up with the 1080ti ;)

1585751236884.png
 
I just looked up some 8 Ghz GDDR6 on digikey for curiosity's sake. These appear to be 512 Mb ICs vs say the 1Gb on a 2080 Ti? so no idea how that would impact the price, but still pretty cheap compared to the numbers I've seen floating around above, and give you an idea about bulk pricing when the minimum quantity is 2000.
View attachment 234682

Huh? Those numbers are stupid high. WAY WAY WAY higher than talked about in this thread. The 2080Ti has 11 memory chips on it, even assuming those 512MB prices that would be $269 in memory cost. We know Nvidia and AMD pay less than those costs, but what you showed there is ridiculously expensive.

In order to go above 12GB, Nvidia either has to use some 2GB chips which cost more than double the price of a 1GB chip or they have to use two 1GB chips per bus interface which means you cant actually access all the memory at once and must switch between chips for transactions which slows the memory down for having that much. Either way to raise the memory to stupid numbers like 22GB or even 32GB on some other model cards would cost about double what they are paying now, and would result in no better performance and may actually result in worse performance.
 
Last edited:
Huh? Those numbers are stupid high. WAY WAY WAY higher than talked about in this thread. The 2080Ti has 11 memory chips on it, even assuming those 512MB prices that would be $269 in memory cost. We know Nvidia and AMD pay less than those costs, but what you showed there is ridiculously expensive.

In order to go above 12GB, Nvidia either has to use some 2GB chips which cost more than double the price of a 1GB chip or they have to use two 1GB chips per bus interface which means you cant actually access all the memory at once and must switch between chips for transactions which slows the memory down for having that much. Either way to raise the memory to stupid numbers like 22GB or even 32GB on some cards would cost about double what they are paying now, and would result in no better performance and may actually result in worse performance.

Then how do they get 48GB of GDDR6 on the Quadro?
 
Then how do they get 48GB of GDDR6 on the Quadro?
The same way you would get 22GB on a 2080Ti, 11 on the front side 11 on the back side. But this is connecting 2 chips to each bus, and you cannot access both chips in the same transaction so while you have more memory it isnt exactly the best configuration. These situations the manufacturer has to weight the benefit of more ram vs faster ram for the target audience, and workstation tends to need bigger ram above all because if you can fit your whole render job in memory then it will be way faster than if you cannot.

11 chips or 22 chips would be for the same 352-bit bus of the 2080Ti. 384-bit would be 12 or 24 chips. 24 chips, at 2GB each = 48GB.
 
The same way you would get 22GB on a 2080Ti, 11 on the front side 11 on the back side. But this is connecting 2 chips to each bus, and you cannot access both chips in the same transaction so while you have more memory it isnt exactly the best configuration. These situations the manufacturer has to weight the benefit of more ram vs faster ram for the target audience, and workstation tends to need bigger ram above all because if you can fit your whole render job in memory then it will be way faster than if you cannot.

11 chips or 22 chips would be for the same 352-bit bus of the 2080Ti. 384-bit would be 12 or 24 chips. 24 chips, at 2GB each = 48GB.

So they need to redesign the entire pcb to really toss more ram on the card or use a sandwich config like this:

Using some super speed interconnect. Place all the vram on a separate card. Latency maybe lots more?
 

Attachments

  • SmartSelect_20200401-134836_Firefox.jpg
    SmartSelect_20200401-134836_Firefox.jpg
    267.2 KB · Views: 0
I just looked up some 8 Ghz GDDR6 on digikey for curiosity's sake. These appear to be 512 Mb ICs vs say the 1Gb on a 2080 Ti? so no idea how that would impact the price, but still pretty cheap compared to the numbers I've seen floating around above, and give you an idea about bulk pricing when the minimum quantity is 2000.
View attachment 234682
GB, not Gb. If that is Gb then it would cost $4,205.03 to put 11GB of VRAM on the 2080 Ti.
The same way you would get 22GB on a 2080Ti, 11 on the front side 11 on the back side. But this is connecting 2 chips to each bus, and you cannot access both chips in the same transaction so while you have more memory it isnt exactly the best configuration. These situations the manufacturer has to weight the benefit of more ram vs faster ram for the target audience, and workstation tends to need bigger ram above all because if you can fit your whole render job in memory then it will be way faster than if you cannot.

11 chips or 22 chips would be for the same 352-bit bus of the 2080Ti. 384-bit would be 12 or 24 chips. 24 chips, at 2GB each = 48GB.
There are 16Gb GDDR6 chips being made, so you could still get away with 11 for 22GB.
 
Back
Top