NVIDIA GeForce RTX 5070 SUPER Possible Specs Emerge: 24 Gbit Memory Takes Centerstage

[btarunr]

NVIDIA will tap into 24 Gbit (3 GB) GDDR7 memory chips and increase memory sizes across the board for its RTX 50-series SUPER line of GPUs slated for later this year, if leaked specs of the upcoming GeForce RTX 5070 SUPER are anything to go by. Kopite7kimi, a reliable source with NVIDIA leaks, says that the RTX 5070 SUPER will feature 18 GB of memory—that's six 24 Gbit memory chips across a 192-bit wide GDDR7 memory bus. The card has a SKU designation of "PG147-SKU65" and ASIC code of "GB205-400-A1."

Besides 18 GB of memory, the RTX 5070 SUPER reportedly maxes out the "GB205" silicon it's based on, enabling all 50 SM present. The current RTX 5070 has 48 out of 50 SM enabled for 6,144 CUDA cores, whereas the RTX 5070 SUPER, with its 50 SM, should have 6,400 CUDA cores. The marginal increase in shader count will be bolstered by the 50% increase in memory size, and possible increases in GPU boost speeds. The memory bandwidth, however, remains unchanged. It ticks at the same 28 Gbps as on the RTX 5070, which yields 672 GB/s. The TGP will be increased to 275 W, from 250 W.



View at TechPowerUp Main Site | Source

 

[ixi]

Cool, cool, anyway 9070 xt

 

[Prima.Vera]

Basically they just confirmed the 5080 Super also, considering the 18GB VRAM for a 5070 card.

 

[ARF]

This seems a change in the Nvidia strategy. Previously, they always underdelivered VRAM, now they feel the shifting market.
The problem remains - they have no know-how to move to a post-5nm process node. And that will kill their business.

 

[Pumper]

I guess I'd take one for under 600€.

 

[Daven]

Videocardz is hypothesizing the specs of all super variants.

https://videocardz.com/newz/nvidia-also-planning-geforce-rtx-5070-ti-super-with-24gb-gddr7-memory

If true, we are only looking at game and app performance scenarios that benefit from more memory. Everything else is mostly the same.

 

[dj-electric]

This seems a change in the Nvidia strategy. Previously, they always underdelivered VRAM, now they feel the shifting market.
The problem remains - they have no know-how to move to a post-5nm process node. And that will kill their business.

You honestly believe that?
NVIDIA is maximizing profits on a cheaper fabrication node - why would they push themselves to pay more for silicon they can sell for top dollar anyway?
Don't worry for NVIDIA and their giant army of engineers, if they need to go more advance they will.

Regarding that shift. I don't see it yet. I don't see it in the poor RTX 5060 and Ti 8GB models that were just launched.

 

[RejZoR]

This seems a change in the Nvidia strategy. Previously, they always underdelivered VRAM, now they feel the shifting market.
The problem remains - they have no know-how to move to a post-5nm process node. And that will kill their business.

They made everyone buy underpowered GPU's with too little VRAM and now they'll sell you another same chip with more VRAM. So, anyway, I bought Radeon because screw NVIDIA.

 

[Grimling]

but still 5nm process, while we (tsmc) are already @2nm mass production.

 

[Onasi]

Honestly, I will believe it when I see it. So far seems like it obvious clickbait rumors - just maxing out the GB205 and stacking new memory modules. Amazing, what nobody seemingly questions is how much faster would just two additional SMs and a power bump really would be. 5% maybe? The VRAM addition would be mostly irrelevant for this card, but sure, at least it would placate the edge case fanatics.

but still 5nm process, while we (tsmc) are already @2nm mass production.

Tell me you don’t understand silicon design and manufacturing without outright doing so.

 

[Macro Device]

50 SM feels extra low effort and it only enables being roughly in-between 9070 and 9070 XT in terms of raw speed but... Do they actually need to offer more than that in this price range?

5070 TS feels nasty. It'll be the cheapest 24 GB CUDA card out here and I'm sure as hell people will hunt it down as a modern 3090 / cheaper 4090 replacement so the price keeps above $1000.

5% maybe?

Outta box, rather 10% because the clocks are gonna be higher by about additional 3 or maybe even 5 percent. But yes, that means only possible to tell in a blind test if the 12 gig version is VRAM-starved, otherwise they're almost identical.

 

[close]

This seems a change in the Nvidia strategy. Previously, they always underdelivered VRAM, now they feel the shifting market.
The problem remains - they have no know-how to move to a post-5nm process node. And that will kill their business.

TSMC allocates (some?) capacity on the cutting edge nodes via bidding process. At lest what's left after Apple takes the lion's share. Nvidia has zero incentive to overpay just to be on the latest node as long as they sell just fine even on an older, cheaper node. Since we're talking about consumer GPUs, Nvidia has over 90% of the market *and growing*. Fabrication node just isn't a problem for them, especially when it's a choice.

When this becomes a true bottleneck and hinders their growth I'm sure the ~$4tn company will find a way to fit a shrink into their design and budget. Manufacturing, which is the toughest nut to crack (just ask Intel) is already a solved problem by TSMC so Nvidia doesn't have to worry about that.

 

[Tigerfox]

It might be finally on par with 9070 and 4070S then, but GM205 is too much cut down to be good midrange GPU and 18GB might make it to expensive.

I fear there might be no 5070Ti 24GB, though.

50 SM feels extra low effort and it only enables being roughly in-between 9070 and 9070 XT in terms of raw speed but

GM205 simply doesn't have more cuda cores, so for more theay would have to use GM203, which is much more expensive. AFAIK 5070 is slightly below 9070, so 5070S might be slightly above or on par, but nowhere near 9070XT.

Outta box, rather 10% because the clocks are gonna be higher by about additional 3 or maybe even 5 percent.

Don't bet on it, the higher densitiy VRAM an the additional SM will eat up most of the additional TDP.

 

[Dirt Chip]

Go Go 5060\Ti\SUPER 12Gb

 

[SOAREVERSOR]

This seems a change in the Nvidia strategy. Previously, they always underdelivered VRAM, now they feel the shifting market.
The problem remains - they have no know-how to move to a post-5nm process node. And that will kill their business.

This will not kill their business.

What's nvidia's actual product? It's CUDA. When it comes to GPU compute nvidia is effectively a monopoly. When it comes to mUh G4m1ngPC!!!! PC gamers are the stupidest brand chasing customers on the planet. Everyone will keep buying nvidia and hope nvidias actions and AMDs wins lead to cheaper nvidia and when it does not they will still buy nvidia.

Complaining about what they deliver doesn't work because everyone will buy it regardless. And if we are truly honest about PC gaming it's vastly inferior to console gaming. Most PC gaming is games that do not consume that many resources and even though the most common monitor resolution may say 1080p many people with those monitors are in regions where they run at 720p or strip out the details anyways.

Then there is the issue that nvidia is firmly aware that PC gaming is moving to gaming as a service in the cloud. You are going to be renting a GPU with different prices depending on the resolution, frame rates, and details you won't. Cost going up per tier. Everything going on right now is a stop gap to get there.

 

[duckface]

Companies should focus on VRAM instead of just increasing graphics processing capacity. I think we should already have cards with 1TB VRAM for the market. Today, with AI, the more space to store AI models in VRAM, the better. Even more so with multimodal models that do several things at the same time. Home consumers are already getting stuck in VRAM with open source models of 20+ GB being released.

 

[N/A]

The 24Gbit memory also has a nominal speed of 40 Gbps, 28 is limited on purpose., so I'm waiting for the 6070 instead. Especially when N3 production is located in the same fab.

 

[Hyderz]

perhaps next year we might see a new and improved frame gen? or entirely something new that requires additional vram....
but looking at the performance of current 50 series the jump to super models isnt a big leap from the looks of it besides the vram bump...

if it indeed is something new that nvidia is bringing... better to wait for 60 series.. i believe thats where we will see bigger gains

 

[Macro Device]

Don't bet on it, the higher densitiy VRAM an the additional SM will eat up most of the additional TDP.

Like @N/A mentioned above, this new GDDR7 is rated for higher speeds and I might also add it's more power efficient. The VRAM will eat LESS compared to non-Super variants. Hence my prediction of +10% rather than +5% you woulda expected from just +2 SM. 5070 has probably the craziest OC potential in the whole line-up so I won't be surprised to see 5070S running 3+ GHz outta box.

AFAIK 5070 is slightly below 9070, so 5070S might be slightly above or on par, but nowhere near 9070XT.

You do realise 9070 and XT are like 10 percent apart? It's too close to call. I consider them effectively the same GPU.

 

[ARF]

This will not kill their business.

The thing is that both the 3nm and 2nm wafers will become more expensive from now on, not cheaper. And since Nvidia hasn't found reasons to shift the production to the new processes, I doubt they will in the near future.
Guess what - if RTX 5000 rebrand is scheduled for next year, then when will RTX 6000 enter production? In 2028?

You do realise there is physics, you can't shrink the transistors any longer, you don't offer generational improvements. And now explain - the ever poorer average people, how ill they afford more of the same for so many years to come ?

Moore's law is dead.

 

[RejZoR]

Companies should focus on VRAM instead of just increasing graphics processing capacity. I think we should already have cards with 1TB VRAM for the market. Today, with AI, the more space to store AI models in VRAM, the better. Even more so with multimodal models that do several things at the same time. Home consumers are already getting stuck in VRAM with open source models of 20+ GB being released.

They just need to add sufficient amount of VRAM. Not 8GB bs on low end ones or high end ones with barely enough.

 

[phints]

18GB is great for 5070S, even though it remains at 192-bit bus it solves the VRAM issue for a while. Problem now is power, it just keeps going up. 4070 used 200W now up to 275W at same tier. Nvidia desperately needs TSMC 3N.

 

[yfn_ratchet]

The extra VRAM makes it a done deal between it and a 9070 for most, though knowing Nvidia they'll slap an extra $70 on this card for those 24Gb chips. Sigh... another day I longingly look at a 9060XT and wonder about what could have been.

 

[chrcoluk]

5070 has only been out 5 minutes, they need to slow down.

18gigs? Nvidia continues to have weird memory capacities.

 

[lexluthermiester]

I wonder if this means there will be 12GB versions of the 5060 & 5060ti? IF they do this and they keep the price in check, they'll fill hole in the budget market and shut the nay-sayers up.