MSI Typos Hint at 24GB Variant of RTX 5080

[btarunr]

MSI made at least two typos and corrections spaced apart from each other that hint at the possibility of a 24 GB variant of the NVIDIA GeForce RTX 5080. The first instance was on the side of the retail box of an RTX 5080 Vanguard OC Launch Edition, which had "24 GB GDDR7" printed on top of the RTX 5080 label. The second instance was in the VGA Support list in the product page of the MAG X870 Tomahawk motherboard, where the memory size is displayed as 24,576 MB (24 GB). There are two possible interpretations of this.

One possibility is that NVIDIA originally planned the RTX 5080 to feature 24 GB of memory, using 3 GB (24 Gbit) GDDR7 memory chips across the 256-bit wide memory bus of the "GB203" silicon, and initially communicated this to its board partners, but revised the specs in the months leading up to the launch, with the card ending up with 16 GB (8x 16 Gbit). The other possibility is that there is actually a 24 GB variant in the works, which NVIDIA could end up calling the "RTX 5080 24 GB." The RTX 5080 in the market already maxes out the "GB203" silicon it's based on, so the variant could just offer a 50% increase in memory size at the same or similar memory bandwidth. The additional memory could come in handy with 4K Ultra HD + ray tracing + DLSS 4 MFG scenarios where a gamer is looking to turn their roughly 60 FPS framerates to match the 120 Hz refresh rates of their displays.



Then there's SUPER. NVIDIA possibly decided against 24 GB so it could launch such a card as the "RTX 5080 SUPER" without having to tap into the larger "GB202" chip. Since it already maxed out the "GB202," the RTX 5080 SUPER could come with 24 GB, possibly higher memory speeds, such as 32 Gbps, and higher GPU clock speeds. NVIDIA could offer up the SKU at a price-point similar to that of the current RTX 5080—$999. This would probably prove difficult to do with the coaster-sized "GB202."

View at TechPowerUp Main Site | Source

 

[Fishymachine]

The 3GB dies run at 36 and 42 GT, so regardless of vendor they should also come with a bandwidth bump

 

[3DVCash]

If this really is the 5080 "Super" it would HAVE to hit the $999 price point to move the needle at all. At current 5080 prices, this would be almost a 4090 for almost the same price over 3 years later.

 

[Battler624]

The 3GB dies run at 36 and 42 GT, so regardless of vendor they should also come with a bandwidth bump

And nvidia will most likely downclock them or call it a 5080Super or ti if they wanna keep the increased bandwidth.

 

[docnorth]

MSI insist on this, could be a hint instead of a typo. But a 5080 super with the same chip and only memory quantity and bandwidth increase won't be a big upgrade (again).

 

[Prima.Vera]

If there is another Super and Ti in the works, then there is absolutely no reason at all buying the current overscalped RTX 5080. NONE!

 

[Hyderz]

If there is another Super and Ti in the works, then there is absolutely no reason at all buying the current overscalped RTX 5080. NONE!

Then when the 5080 super releases it will be faster by 5fps and achieves 200+fps with frame gen and price will be the same

 

[Veseleil]

This is for the LLM sake, nothing else.

 

[RayneYoruka]

Well well well a worthy rumour of myself. Now we shall see it's simply a rumour or not.

 

[qlum]

This could work out quite well for nvidia, give it a price bump and have dies that are more profitable to sell than the larger die needed for 5090's. While offering the ml crowd a less bad deal in the process.

 

[Gameslove]

Think true! History of the MSI GT 760 4 GB VRAM.

 

[bug]

I imagine multiple versions of any SKU are evaluated (some of them even built) to figure out ROI, bottlenecks and such.

 

[mama]

Another overpriced, unavailable card. Yippee.

 

[EatingDirt]

A 24GB 5080 Super/Ti makes sense if there are enough 5090's that can't be a 5090. Might as well do something useful with those dies, though I see it more likely that they'll be turned into workstation cards rather than retail cards, unless they have an absolutely huge amount of bad 5090's.

There's no real need for a 5080 Super similar to the previous generation 4080 Super(~1% better performance), lower MSRP, as they have no competition at that performance level this time around.

 

[L0stS0ul]

Finally! They finally showed the true version of the RTX 5070. Because the results of the RTX 5080 correspond to what the RTX 5070 should present. 10% increase is rather a joke on gamers and consumers -
Nobody can take the results of the 5080 seriously. The RTX 5080 should catch up or overtake the RTX 4090 - especially if such a tiny card cost 1000$ MSRP and in reality it is 1200$ and more.

 

[GhostRyder]

I would more bet this is going to be the RTX 5080ti (unless they rename it, but still same premise). I doubt they are just going to release an RTX 5080 with more memory.

 

[igormp]

A 24GB 5080 Super/Ti makes sense if there are enough 5090's that can't be a 5090. Might as well do something useful with those dies, though I see it more likely that they'll be turned into workstation cards rather than retail cards, unless they have an absolutely huge amount of bad 5090's.

There's no real need for a 5080 Super similar to the previous generation 4080 Super(~1% better performance), lower MSRP, as they have no competition at that performance level this time around.

You seem to be assuming they would make a 5080Super/Ti based on the GB202 from the 5090, but the talks from the OP are more about using the existing GB203 with the exact core counts of the current 5080, but with bigger and faster memory modules.

 

[Knight47]

5080 24GB was the initial plan most likely, but they couldn't get the 3GB dies ready because they rushed to get it out before tariffs

 

[forman313]

RTX 5080 24GB is very real AFAIK. Nvidia calls it RTX 5090 Mobile. Granted, the 5090 Mobile has 2% less SU´/SM´s.

Taking flawed 5080 dies, massive down clock voila, RTX 5090 Mobile.. The 4070 performance is irrelevant. If they say its a 5090, its a 5090.

5080 Super is probably going to be the 24GB version with 10% higher clocks and 15-20% higher TDP.

 

[Ruru]

I doubt that there will be a Super or Ti since the 5080 has already the full chip. Could be just called as 5080 24GB.

 

[10tothemin9volts]

This denser/higher VRAM consumer GPU for local AI LLM self-hosting indeed is nice, but I wish the 3GB chips were on the much cheaper 5070 Ti. The 5080 is only 13% (1440p) to 15% (4K) faster than the 5070 Ti, but costs 33% more (MSRP).

RTX 5070 Ti	750 $	256-bit	8 GDDR7 2GB chips (32bit per chip)	16 GB	896 GB/s	47 $ per GB
RTX 5080	1000 $	256-bit	8 GDDR7 2GB chips (32bit per chip)	16 GB	960 GB/s	63 $ per GB

Obviously, using the 5080 instead of the cheaper 5070 Ti as a base, is a way for companies to up-sell the cheap silicon. The 5070 Ti may follow later..gota rip off the customers and maximize the profits first.

 

[phints]

Useless for gaming for the next few years at least, but neat I guess.

 

[bug]

Useless for gaming for the next few years at least, but neat I guess.

All we know is somebody toyed with the idea. We don't know if it was deemed worthy. That said, I can see increased VRAM demand from those that would run LLMs locally. At the same time, I see AMD and Nvidia pushing those users towards their professional line of cards, rather than messing with their consumer products lineup.

 

[igormp]

We don't know if it was deemed worthy. That said, I can see increased VRAM demand from those that would run LLMs locally.

There are datacenters in china offering those 4090 48GBs already haha
There are even some of those 4090 48GB on vast.ai already: