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

[Avro Arrow]

Wow, nVidia is actually going to give their "high-rollers" adequate VRAM? It kinda makes you wonder how much it's gonna cost though, eh? There is a sentence from the article that alludes to something but is ultimately devoid of information (it's not the author's fault, he's only going by what the tweet said):

"The key morsels from that Tweet/X were that the purported GeForce RTX 4080 Super (maybe Ti) would arrive in early 2024. Moreover, it would be based on the AD102 GPU, have a TGP of under 450W, yet be priced similarly to the existing RTX 4080."

Whether or not one price is similar to another is a matter of opinion based on what criteria a person has for qualifying something as "similar". We've seen that nVidia has this fetish for charging an extra 25% (or $100) when adding VRAM to a card. Some might say that nVidia charges $100 for 8GB extra VRAM but let's be honest here, you're deluding yourself if you think that nVidia will only charge an extra $50 here.

We've seen from Intel that an extra 8GB of VRAM costs an extra $20 when looking at the MSRP of the A770 8GB and A770 16GB while nVidia increased the MSRP of the already-bad-value RTX 4060 Ti 8GB by $100 for the 16GB version. I have no belief that the RTX 4080 Super will cost less than $100 more than the RTX 4080.

 

[palladin9479]

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

Read before you post.

I know it'll slow down your posting speed and you wont' get in as many each day, but it would vastly improve the quality.

 

[bit_user]

Read before you post.

I know it'll slow down your posting speed and you wont' get in as many each day, but it would vastly improve the quality.

You edited the post after I replied. I quoted the entire content of your post, at the time.

 

[palladin9479]

You edited the post after I replied. I quoted the entire content of your post, at the time.

The edit was before you posted.

 

[bit_user]

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.

I assume 2 DRAM chips per channel works just like having a dual-rank DDR DIMM.

According to Nvidia, the memory bandwidth of the RTX 6000 is 960 GB/s. Compared to the RTX 4090's 1008 GB/s, that's not much of a deficit.

​

https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/rtx-6000/proviz-print-rtx6000-datasheet-web-2504660.pdf

​

 

[bit_user]

The edit was before you posted.

You're looking at the timestamp of when I hit "post" - not when I hit "reply". The quote gets taken when the user hits the "reply" button. It took me some time to compose my message, because I actually fact-checked what I posted.

 

[palladin9479]

I assume 2 DRAM chips per channel works just like having a dual-rank DDR DIMM.

According to Nvidia, the memory bandwidth of the RTX 6000 is 960 GB/s. Compared to the RTX 4090's 1006 GB/s, that's not much of a deficit.

​

https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/rtx-6000/proviz-print-rtx6000-datasheet-web-2504660.pdf

​

More complicated, system DRAM is specifically designed to work in serial, GPU DRAM is tuned to work single chip to bus. It's all possible obviously but has severe consequences for overall performance, but that would depend on workload. When we talk GPU's we're normally talking 3D rendering and wanting to get from CPU decision to GPU compute to screen drawing ASAP, so all the tuning is around that goal. If the goal is less time sensitive and more quality orientated, like the workstation cards your referencing, then we can accept an even more sluggish memory subsystem in exchange for doubling the capacity.

The article was about a new variant of 4080 gaming card, so the context of all discussions is centered on that.

 

[George³]

Read before you post.

I know it'll slow down your posting speed and you wont' get in as many each day, but it would vastly improve the quality.

Based on the same Ada Lovelace architecture and with the same core specs as the RTX 4060 Ti Founders Edition, the sole difference is the use of two 2GB memory chips on each memory channel, doubling the capacity to 16GB

Review. Not needs special controller. No needs more channels and lines. Just soldering chips on other side of PCB and maybe tweak drivers.

 

[Avro Arrow]

No.

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

It's simple. The public has demonstrated that it's willing to pay whatever for them.

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

And yet, the RTX 4080 did... 😆

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.

It's not about how much VRAM a card needs, it's about how little they can get away with and still have the clueless noobs buy them. They're a business and they're in the business of raking in money. How they accomplish that is incidental.

Consoles aren't exactly cheap either

No, but they cost way less than a lot of video cards.

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.

The RX 7900 XTX will outperform it simply because it has more GPU horsepower. The 256-bit bus of the RTX 4080, while inferior to the 384-bit bus of the RX 7900 XTX, won't have a big effect on the addition of only 4GB of extra VRAM. It's not true that nVidia handicapped the RTX 4060 Ti 16GB, they just didn't increase the bandwidth from the 8GB version.

The RTX 4060 Ti's fram buffer was increased by 8GB on a 128-bit bus and that could cause problems (it didn't really cause any major problems though) but in this case, you're increasing the VRAM by half of 8GB with a bus that's already double that of the RTX 4060 Ti.

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.

Yep, and those RX 7600 XT 10GB rumours don't seem to be going away so they may have another player in that space.

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

I look at it this way...
Intel Arc A770 8GB MSRP = $329USD
Intel Arc A770 16GB MSRP = $349USD

Intel added 8GB of VRAM to the A770 and proceeded to charge an extra $20USD for it. Since no corporation does anything for free, this means that it actually cost Intel LESS than $20 per card for the 8GB VRAM increase. There's no way that nVidia would pay significantly more for VRAM than Intel (or AMD for that matter) since, as the overwhelming market leader, they're obviously the biggest single VRAM customer in the world for the companies that produce it. That means, if anything, nVidia pays even less than Intel or AMD for VRAM.

This alone demonstrates that adding 8GB of VRAM to a card is neither difficult nor expensive and anyone who tries to claim otherwise is either a blind and clueless nVidia fanboy or actually works for nVidia.

When someone's words contradict something that has actually happened already (in this case, the tiny price disparity between the 8 and 16GB versions of the Intel Arc A770), they're either stupid, lying or both. It's just that simple.

 

[bit_user]

GPU DRAM is tuned to work single chip to bus.

According to whom? From what I'm reading, it turns out that recent GDDR memories have a specific feature to enable 2 chips per channel.

"DDR5 and GDDR5X SGRAMs support a x16 mode in which two devices are assembled on the top and bottom on the PCB in a clamshell configuration. This x16 mode is an elegant way to double the density per memory channel as seen by the memory controller (host), as the two devices are connected to the same command/address bus and therefore always work in sync. The 32-bit data bus from the memory controller is split, with 16 bits routed point-to-point (P2P) to the top and bottom devices. This x16 mode is configured at boot time.

GDDR6 supports the exact same configuration option, except the name has changed from x16 mode to x8 mode, following the dual-channel architecture of GDDR6. In x8 mode, only one of the two data bytes per channel is enabled (byte 0 of channel A and byte 1 of channel B), while the other two data bytes are disabled. GDDR6 does not require a mirror function (MF) pin for this purpose, thanks to the dual-channel architec-
ture.

The figure below illustrates how two GDDR6 devices in x8 mode are connected to give the memory controller the same view as a single GDDR6 device in normal (x16) mode: <figure omitted - see source link>

• For memory channel A, byte 0 comes from the top device and byte 1 from the bottom device.
• For memory channel B, byte 0 comes from the bottom device and byte 1 from the top device.

The bottom device is flipped over and the two bytes of that device are logically swapped between the two channels. Channel A and B command/address pins are routed point-to-two-point (P22P) to both devices, preferably by simple vias.

Source: https://www.micron.com/-/media/client/global/documents/products/technical-note/dram/tned03_gddr6.pdf
​

 

[palladin9479]

Review. Not needs special controller. No needs more channels and lines. Just soldering chips on other side of PCB and maybe tweak drivers.

Better break down

NVIDIA GeForce RTX 4060 Ti 16 GB Review - Twice the VRAM Making a Difference?

Finally! Our GeForce RTX 4060 Ti 16 GB review is live. We're investigating what performance gains you can expect from doubling the VRAM size to 16 GB. Our review of the 16 GB version also looks at power consumption and how overclocking is affected by the additional four memory chips on the back...

www.techpowerup.com

Soldiered them back to back to minimize the extra length on the bus, then use better quality DRAM chips.

According to whom?

Physics.

Specifically the longer the distance a signal needs to travel the more time it needs to get there and the more time you have to wait until you can send the next signal. It's the exact same problem that crops up with using four sticks of memory instead of two sticks of memory, and why we insist on using matched sets.

 

[George³]

Better break down

Yes overclockers are screwed. That is irrelevant to this discussion. In fact, we've all deviated anyway, though not very much.

 

[palladin9479]

Yes overclockers are screwed. That is irrelevant to this discussion. In fact, we've all deviated anyway, though not very much.

My reponsce

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.

It's why GPU manufacturers don't do it, unless for a very specific reason. Complexity is the enemy of stability. In the world of IT nearly anything is possible if someone is willing to try hard enough or find a niche use case. Doesn't mean it's a good idea, especially towards market center mass. When nVidia released that 16GB version for $100 more, everyone screamed at the "price gouging", and while nVidia is definitely ... rather aggressive with pricing, there are real additional costs to making that product for no real benefit.

 

[bit_user]

Physics.

Specifically the longer the distance a signal needs to travel the more time it needs to get there and the more time you have to wait until you can send the next signal.

According to physics, the speed of electrical signal propagation in copper is (conservatively) about 150 mm per ns. According to data, the memory latency of the RTX 4090 is about 269 ns. A signal could easily traverse 40 meters of copper, in that time.

Source: https://chipsandcheese.com/2022/11/02/microbenchmarking-nvidias-rtx-4090/
​

So, by that metric, adding a couple mm's for the "simple vias" described above should have an immeasurable impact on latency.

In fact, I think the impact is so minimal that the reason the RTX 6000 only has 960 GB/s of memory bandwidth isn't due to the additional latency, but perhaps due to greater electrical load of driving the extra chips, or perhaps simple reliability concerns. Server parts are often de-rated so they stay more comfortably within operating margins.

It's the exact same problem that crops up with using four sticks of memory instead of two sticks of memory, and why we insist on using matched sets.

What we're talking about here is akin to single-sided DIMMs vs. double-sided ones. The path length to each set of chips is virtually identical, which is why there's not generally a clock penalty for using dual-rank DIMMs. Single-rank DIMMs reportedly overclock better, but that's likely more to do with having a lower electrical load.

 

[George³]

My reponsce

Your response i refuse to comment because length of way of signal just is...same. Why you think that is double.

 

[Papusan]

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.

Only rumors but nvidia try everything they can to keep up the prices...

View: https://youtu.be/zcJZrWS5xfc?t=511

 

[Geezer760]

It's simple. The public has demonstrated that it's willing to pay whatever for them.

And yet, the RTX 4080 did... 😆

It's not about how much VRAM a card needs, it's about how little they can get away with and still have the clueless noobs buy them. They're a business and they're in the business of raking in money. How they accomplish that is incidental.

No, but they cost way less than a lot of video cards.

The RX 7900 XTX will outperform it simply because it has more GPU horsepower. The 256-bit bus of the RTX 4080, while inferior to the 384-bit bus of the RX 7900 XTX, won't have a big effect on the addition of only 4GB of extra VRAM. It's not true that nVidia handicapped the RTX 4060 Ti 16GB, they just didn't increase the bandwidth from the 8GB version.

The RTX 4060 Ti's fram buffer was increased by 8GB on a 128-bit bus and that could cause problems (it didn't really cause any major problems though) but in this case, you're increasing the VRAM by half of 8GB with a bus that's already double that of the RTX 4060 Ti.

Yep, and those RX 7600 XT 10GB rumours don't seem to be going away so they may have another player in that space.

I look at it this way...
Intel Arc A770 8GB MSRP = $329USD
Intel Arc A770 16GB MSRP = $349USD

Intel added 8GB of VRAM to the A770 and proceeded to charge an extra $20USD for it. Since no corporation does anything for free, this means that it actually cost Intel LESS than $20 per card for the 8GB VRAM increase. There's no way that nVidia would pay significantly more for VRAM than Intel (or AMD for that matter) since, as the overwhelming market leader, they're obviously the biggest single VRAM customer in the world for the companies that produce it. That means, if anything, nVidia pays even less than Intel or AMD for VRAM.

This alone demonstrates that adding 8GB of VRAM to a card is neither difficult nor expensive and anyone who tries to claim otherwise is either a blind and clueless nVidia fanboy or actually works for nVidia.

When someone's words contradict something that has actually happened already (in this case, the tiny price disparity between the 8 and 16GB versions of the Intel Arc A770), they're either stupid, lying or both. It's just that simple.

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.

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.

Right, what you said (nVidia deliberately hamstring their entire product line to upsell).

 

[Order 66]

Right, what you said (nVidia deliberately hamstring their entire product line to upsell).

That is basically what I (along with the others you quoted) have been saying the whole time. It sure is nice to have someone agree with my sentiment rather than arguing pointlessly about it. I saw someone where their argument for why games were so demanding was because they were "state of the art" not because they were unoptimized.

 

[bit_user]

It sure is nice to have someone agree with my sentiment rather than arguing pointlessly about it.

If I agree with someone's post and have nothing substantial to add, I just click its "Like" button. I tend to be pretty generous with likes.

 

[helper800]

If I agree with someone's post and have nothing substantial to add, I just click its "Like" button. I tend to be pretty generous with likes.

Same here on both accounts.

 

[Order 66]

If I agree with someone's post and have nothing substantial to add, I just click its "Like" button. I tend to be pretty generous with likes.

I tend to be rather generous with likes as well, a lot of times it seems like a way to not only agree with someone's post but also let them know that you have seen it. I usually assume (though this is not always the case) that when someone likes one of my posts, they will also quote it and add their own opinions.

 

[Order 66]

It's simple. The public has demonstrated that it's willing to pay whatever for them.

And yet, the RTX 4080 did... 😆

It's not about how much VRAM a card needs, it's about how little they can get away with and still have the clueless noobs buy them. They're a business and they're in the business of raking in money. How they accomplish that is incidental.

No, but they cost way less than a lot of video cards.

The RX 7900 XTX will outperform it simply because it has more GPU horsepower. The 256-bit bus of the RTX 4080, while inferior to the 384-bit bus of the RX 7900 XTX, won't have a big effect on the addition of only 4GB of extra VRAM. It's not true that nVidia handicapped the RTX 4060 Ti 16GB, they just didn't increase the bandwidth from the 8GB version.

The RTX 4060 Ti's fram buffer was increased by 8GB on a 128-bit bus and that could cause problems (it didn't really cause any major problems though) but in this case, you're increasing the VRAM by half of 8GB with a bus that's already double that of the RTX 4060 Ti.

Yep, and those RX 7600 XT 10GB rumours don't seem to be going away so they may have another player in that space.

I look at it this way...
Intel Arc A770 8GB MSRP = $329USD
Intel Arc A770 16GB MSRP = $349USD

Intel added 8GB of VRAM to the A770 and proceeded to charge an extra $20USD for it. Since no corporation does anything for free, this means that it actually cost Intel LESS than $20 per card for the 8GB VRAM increase. There's no way that nVidia would pay significantly more for VRAM than Intel (or AMD for that matter) since, as the overwhelming market leader, they're obviously the biggest single VRAM customer in the world for the companies that produce it. That means, if anything, nVidia pays even less than Intel or AMD for VRAM.

This alone demonstrates that adding 8GB of VRAM to a card is neither difficult nor expensive and anyone who tries to claim otherwise is either a blind and clueless nVidia fanboy or actually works for nVidia.

When someone's words contradict something that has actually happened already (in this case, the tiny price disparity between the 8 and 16GB versions of the Intel Arc A770), they're either stupid, lying or both. It's just that simple.

You make a very well-worded argument, and I agree completely. It's kind of like Apple charging an extra $400 for another 1TB of storage or an extra $800 for 32GB of RAM.