News Nvidia GeForce RTX 4080 Super Rumored to Feature 20 GB VRAM

[Admin]

Nvidia is rumoured to be preparing a beefed-up GeForce RTX 4080 Super with 20 GB VRAM which might help address the China market when RTX 4090 stocks are all gone.

Nvidia GeForce RTX 4080 Super Rumored to Feature 20 GB VRAM : Read more

 

[Papusan]

Nice. So the sanctions against China will determine how much performance you'll get from next gen graphics cards forwards. Great news.

 

[digitalgriffin]

Nice. So the sanctions against China will determine how much performance you'll get from next gen graphics cards forwards. Great news.

No.

I don't understand why 4090 prices are shooting through the roof. Maybe someone can explain it to me.

Let's say 50k Nvidia 4090s were sold to China pre sanctions. That demand is still 50k/month. So why do scalpers and 3rd party resellers shell companies circumventing sanctions think they are going to get rich when the demand is the same? Even if sanctions are worthless and China gets 50k/month then Supply doesn't change. If sanctions work then that's more Supply for the world market. Supply and demand is the same or becomes better.

 

[usertests]

Makes sense, having less than the 7900 XT would be cringe.

 

[Geezer760]

Rumored, rumored? Shouldn't 20gb vram been from the get go, ngreedia has always skimmed (scamming) out on vram, AMD has pretty much always added more vram, even back when I bought my R9 390 (still running today) it had 8gb of vram compared to ngreedia's 6gb vram.

 

[thisisaname]

No.

I don't understand why 4090 prices are shooting through the roof. Maybe someone can explain it to me.

Let's say 50k Nvidia 4090s were sold to China pre sanctions. That demand is still 50k/month. So why do scalpers and 3rd party resellers shell companies circumventing sanctions think they are going to get rich when the demand is the same? Even if sanctions are worthless and China gets 50k/month then Supply doesn't change. If sanctions work then that's more Supply for the world market. Supply and demand is the same or becomes better.

China buying now rather than later would push up demand and thus the price. Nvidia is not going to increase supply of the 4090 and it can sell Ai cards to a massively lot more.

As for scalpers I do not think they are thinking long term, as long as it sell for more than they paid for they do not care.

 

[ingtar33]

the pprice of the 4090 is going up because nvidia stopped production on them a month or two ago. combined with these sanctions.

and the super lineup being contemplated just to get around US export restrictions would be so in line with ngreedia's SOP that I wouldn't be surprised if it's priced identically to the 4090, and released only in china.

 

[palladin9479]

Rumored, rumored? Shouldn't 20gb vram been from the get go, ngreedia has always skimmed (scamming) out on vram, AMD has pretty much always added more vram, even back when I bought my R9 390 (still running today) it had 8gb of vram compared to ngreedia's 6gb vram.

VRAM isn't something you can magically add more of, not without redesigning the chip or creating an asymmetrical memory architecture which is extremely bad for performance.

The only way they could "Add" another 4GB would be if they were also adding another 32-bit memory channel. Currently the 4080 has a 256-bit memory interface, which breaks down into 8 32-bit memory channels. Each memory channel has a single 2GB (16Gb) memory chip. 8 * 2GB = 16GB. The 12GB models all have 192-bit memory bus's with 6 32-bit memory channels. The lowly 8GB model that everyone freaked out about, that's because it has an anemic 128-bit memory interface, four 32-bit memory channels with a 2GB chip on each one. To "Add" 4GB without increasing memory channels would mean one channel had more then the rest causing extremely unpredictable performance problems. The 16GB 4060 also has only 4 32-bit memory channels but use's 4GB (32Gb) memory chips instead, and gets the same performance as the 8GB in almost all tests. You upgrade all the memory chips or none of the memory chips.

Also memory size isn't very important once your past the point of holding all data of the existing scene, memory bandwidth on the other hand is absolutely crucial. What most people think is a memory size issue is actually a memory bandwidth issue, and nVidia deliberately hamstrung their entire product line to upsell people to the 4090.

 

[williamcll]

No.

I don't understand why 4090 prices are shooting through the roof. Maybe someone can explain it to me.

Let's say 50k Nvidia 4090s were sold to China pre sanctions. That demand is still 50k/month. So why do scalpers and 3rd party resellers shell companies circumventing sanctions think they are going to get rich when the demand is the same? Even if sanctions are worthless and China gets 50k/month then Supply doesn't change. If sanctions work then that's more Supply for the world market. Supply and demand is the same or becomes better.

Because AIBs would have a hard time manufacturing them under sanctions. There are very little factories out there that can assemble 4090s for cheap and it's not EVGA.

 

[watzupken]

I am apprehensive here. Looking at RTX 2080 vs 2080 Super, there is no change in the VRAM configuration. I am comparing with the RTX 2080 because it is clear to me that Nvidia decided to go back to this way to segregate their products. RTX 3090/80, made it very difficult to upsell their customers, and they likely don't want to end up in the same situation.
There is also no reason to create a new and more powerful SKU because there is no way they can predict when the product will get sanctioned as well.

 

[lumineZ]

Simple. At the pricing level that is going on overall in the GPU market. I vote with my wallet and I will not buy an 40xx card. Period. And if this keeps up with the next generation. I will most likely be done with PC gaming. Shame since it has been a great and fun hobby since the Commandore64 era.

 

[williamcll]

Simple. At the pricing level that is going on overall in the GPU market. I vote with my wallet and I will not buy an 40xx card. Period. And if this keeps up with the next generation. I will most likely be done with PC gaming. Shame since it has been a great and fun hobby since the Commandore64 era.

Consoles aren't exactly cheap either

 

[bit_user]

VRAM isn't something you can magically add more of, not without redesigning the chip

There are at least four possibilities, here:

1. They refreshed the GPU with more memory channels.
2. The GPU always had more channels that simply went unused.
3. They're using binned AD102 dies.
4. They are actually reducing the 256-bit interface down to 160-bit and using double-density DRAM chips or 2x chips per channel.

I somewhat doubt #1, but it's always possible. #2 also seems unlikely, but I guess it's almost as plausible as the first. #3 seems very plausible, other than I doubt they have that many rejects from RTX 4090 production and those dies can't be cheap to make. #4 is very unlikely, considering the impact on bandwidth and how this is a "Super" card, which are normally faster than the normal ones.

or creating an asymmetrical memory architecture which is extremely bad for performance.

You're probably thinking of the GTX 970, but that was different. That had an extra 8-bit channel, which is why it was so bad.

Having an extra 32-bit channel shouldn't be an issue, especially when you consider that cards like the GTX 1080 Ti were stellar performers, in spite of having the odd memory configuration of 11x 32-bit channels, for a total of 352 bits (hence the 11 GB memory capacity).

Until Nvidia started using such funky memory sizes, I had assumed they interleaved their memory channels. However, I'm not sure about that. It seems to me they could do pretty well by having a linear address mapping and basically interleaving via the page table. That saves you an extra division in your address calculations.

Also memory size isn't very important once your past the point of holding all data of the existing scene, memory bandwidth on the other hand is absolutely crucial. What most people think is a memory size issue is actually a memory bandwidth issue, and nVidia deliberately hamstrung their entire product line to upsell people to the 4090.

You're thinking of gaming, but I suspect the 20 GB upgrade might've been more oriented towards AI workloads. Especially if they had warning signs of the RTX 4090 getting banned from China, but also just to effectively compete with the RTX 7900 XT on LLMs.

 

[williamcll]

There are at least four possibilities, here:

1. They refreshed the GPU with more memory channels.
2. The GPU always had more channels that simply went unused.
3. They're using binned AD102 dies.
4. They are actually reducing the 256-bit interface down to 160-bit and using double-density DRAM chips or 2x chips per channel.

VRAM sizes aren't baked in. It's limited by the space on the PCB. Some cards don't even need a BIOS Change to reconfigure VRAM. Earlier this year there was the Gxore 3070 with double the base VRAM for Digital Signage.

 

[palladin9479]

VRAM sizes aren't baked in. It's limited by the space on the PCB. Some cards don't even need a BIOS Change to reconfigure VRAM. Earlier this year there was the Gxore 3070 with double the base VRAM for Digital Signage.

It's not the PCB, it's the memory bus + chip capacity. Well technically you could be limited on PCB size, but it would have to be some ridiculously massive 768-bit (24 chip) memory interface. Each DRAM chip is a 32-bit data interface that gets routed to a single memory channel. GPU's band multiple memory channels together into what we call a memory bus. The manufacturer can then disable specific channels to sell the chip as a lower end model. The import part here is that number of chips is tied to the size of the bus with the 4080 having only eight chips due to it's 256-bit bus vs the 4090 having 12 via it's 384-bit bus, which is about the biggest nVidia has made so far. AMD played with HBM in it's previous generations, and while that allows for some truly impressive memory bus sizes, it comes with it's own technical challenges.

 

[Order 66]

I bet the 7900xtx will still outperform it becuase even if Nvidia makes a 20GB 4080 super they will still handicap the memory bus like they did with the 4060 ti 16gb.

 

[Order 66]

Simple. At the pricing level that is going on overall in the GPU market. I vote with my wallet and I will not buy an 40xx card. Period. And if this keeps up with the next generation. I will most likely be done with PC gaming. Shame since it has been a great and fun hobby since the Commandore64 era.

I think you forgot AMD exists, at least they haven't completely abandoned the budget segment like nvidia has, I mean the rx 7600 is currently $20 cheaper than the cheapest 4060.

 

[bit_user]

VRAM sizes aren't baked in. It's limited by the space on the PCB. Some cards don't even need a BIOS Change to reconfigure VRAM. Earlier this year there was the Gxore 3070 with double the base VRAM for Digital Signage.

That's a straight doubling. It gets more tricky when you talk about non-2x changes, because those can't be implemented simply by doubling up the number of chips or using higher-density chips.

As for doubling the number of chips - the only reason that's even possible is if it's a PCB that's also used for the workstation cards which have 2x memory capacity. That's usually accomplished by doubling the number of DRAM chips per channel, meaning the PCB will have all the necessary traces for it. It's not like you can just take any old gaming card and necessarily add more DRAM chips to it (other than perhaps upgrading a partially-depopulated GPU, like the GTX 1080 Ti, from 11 GB to 12 GB or maybe even repopulating all its DRAM chips with double-density parts, to reach 24 GB).

 

[bit_user]

I bet the 7900xtx will still outperform it becuase even if Nvidia makes a 20GB 4080 super they will still handicap the memory bus like they did with the 4060 ti 16gb.

If it's made using AD102 dies (i.e. the same die as the RTX 4090 uses), then I doubt it. I just don't know if that's economically viable, or makes sense in light of the current demand for RTX 4090 cards.

 

[Colif]

#3 seems very plausible, other than I doubt they have that many rejects from RTX 4090 production and those dies can't be cheap to make.

It wouldn't be the first time rejects are sold as lesser power versions with disabled parts.

We shall see if the rumour turns real. And how many there are of them.

 

[George³]

That's a straight doubling. It gets more tricky when you talk about non-2x changes, because those can't be implemented simply by doubling up the number of chips or using higher-density chips.

As for doubling the number of chips - the only reason that's even possible is if it's a PCB that's also used for the workstation cards which have 2x memory capacity. That's usually accomplished by doubling the number of DRAM chips per channel, meaning the PCB will have all the necessary traces for it. It's not like you can just take any old gaming card and necessarily add more DRAM chips to it (other than perhaps upgrading a partially-depopulated GPU, like the GTX 1080 Ti, from 11 GB to 12 GB or maybe even repopulating all its DRAM chips with double-density parts, to reach 24 GB).

They just install 2X number of memory chips. In front and on backside.

 

[bit_user]

They just install 2X number of memory chips. In front and on backside.

Yeah, but that only works if the PCB is designed to accommodate it. This is probably true of the Founders Edition and others built using the reference design, since that's likely the same design used to make the workstation cards. For any AIB partners who design their own PCBs, they almost certainly don't have the traces & pads to handle the extra DRAM chips.

 

[Bluoper]

Watching everyone here in the comments fight as to why it is/isn't hard to increase vram is really funny.

 

[palladin9479]

They just install 2X number of memory chips. In front and on backside.

You can't do that, not without also doubling the number of memory channels, which requires redesigning a new GPU. One DRAM chip per 32-bit of memory bus. Well in theory you could serialize the DRAM chips to double capacity, but that has severe consequences for performance timings due to the memory signal having to travel twice as far before being terminated.

 

[bit_user]

You can't do that, not without also doubling the number of memory channels, which requires redesigning a new GPU. One DRAM chip per 32-bit of memory bus.

Bro, seriously. Try looking at the specs of their workstation cards.

The RTX 6000 is based on the same AD102 GPU as the RTX 4090, but it has 48 GB instead of 24 GB. Are there 32 Gb GDDR6 dies out there? No. Did they make a version of the chip with a 768-bit memory interface? Certainly not. They do it by doubling up the number of chips per channel.

So, yes, you can in fact put 2 GDDR chips on the same channel, provided the memory controller is designed to handle the load. You'll probably have to drop the memory speed, but it might be worth it if you actually need the additional capacity. As others have noted, they typically put the extra chips on the backside of the PCB, probably to keep the traces as short & similar as possible.

I didn't find a teardown of the RTX 6000, but here's a A6000 (previous gen) which clearly shows 12 GDDR6 chips on each side: