If the recently rumoured specs for the new Nvidia RTX 5090 and RTX 5080 graphics cards are indeed true we’re looking at a flagship GPU with a terrifying amount of processing power and a second-tier chip that’s already pissing off a lot of gaming folk.
GeForce RTX 5090PG144/145-SKU30GB202-300-A121760FP32512-bit GDDR7 32G600WSeptember 26, 2024
GeForce RTX 5080PG144/147-SKU45GB203-400-A110752FP32256-bit GDDR7 16G400WSeptember 26, 2024
The latest spec ‘leak’ has come from more or less reliable X leaker, kopite7kimi, and tells a story of one almighty RTX 5090, with a specs bump that gives it a mouth-watering amount of TSMC-built tech in a single graphics card. And reportedly still a two-slot reference card at that.
It’s going to make the RTX 4090 look like a clown card in more ways than one. That Founders Edition is a mammoth graphics card that still makes me laugh every single time I slot it into the test rig in the office, such is its vast, novelty-GPU size. That’s going to be going against what is reportedly going to be a dual-slot card will make it look even more outsized. And that’s before you get to talk about the conversely mammoth spec of the reported new GPU. With 21,760 CUDA cores (effective FP32 units)—that’s still not utilising the full GB202 complement of cores—it’s a hefty 33% increase in core count over the RTX 4090, which itself had over 50% more cores than the RTX 3090.
Then the new Nvidia Blackwell flagship GeForce card is reportedly going to be flexing with a 512-bit aggregated memory bus and a full 32 GB of GDDR7 memory. And, guess what, that makes for a 33% increase in both memory and memory bus, too. The increased power numbers (presumably the TGP of the card) suggests a 600 W peak, which, if you’ve been following the maths, also hits that seemingly magic 33% mark.
I wonder what sort of performance improvement Nvidia is going to be touting with its new card? Surely it’s got to do better than 33%? In our testing, the RTX 4090 was, on average, 72% faster than the RTX 3090 in our 4K gaming tests. So, I’m kinda hoping it’s more in line with the memory bandwidth uplift instead.
Because one of the few numbers which doesn’t follow the plus-one-third trend is memory bandwidth the RTX 5090 is supposedly going to have at its disposal. With both that 512-bit bus and the use of 28 Gbps GDDR7, we are looking at 1,792 GB/s in terms of raw memory bandwidth, that’s a pretty incroyable 78% increase over the RTX 4090.
Quite how Nvidia is going to squeeze all that extra power into a dual-slot GPU design, I really don’t know. The rumours are that we’re only looking at a modest change in process node—N4P versus N4 for the Ada generation—so that’s not going to get you running much cooler. So, we could be looking at a very special Founders Edition indeed, potentially some kind of liquid-cooled beastie.
Doubtless the third-party GPU vendors are still likely to be tacking monstrous coolers to their own versions, and I doubt subsequent RTX 50-series Founders Edition cards would get a similar treatment. Though, who knows.
I kinda don’t want to think about how much this thing is going to cost, though. The RTX 4090 was a relative bargain compared to the pricing increases that happened across the rest of the RTX 40-series range, but I struggle to believe that in a world where it has zero competition Nvidia isn’t going to price its flagship GPU anywhere south of $2,000. If we’re going by the 33% numbers around $2,100 makes some sense…
Given that it’s twice the GPU of the rumoured RTX 5080, which is likely to ship at the same $1,000 mark as the RTX 4080 Super, that all kinda follows.
| Header Cell – Column 0 | Nvidia RTX 5090* | Nvidia RTX 5080* | Nvidia RTX 4090 | Nvidia RTX 4080 Super |
|---|---|---|---|---|
| GPU | GB202-300 | GB203-400 | AD102-300 | AD103-400 |
| CUDA Cores | 21,760 | 10752 | 16384 | 10240 |
| SMs | 170 | 84 | 128 | 80 |
| Memory | 32 GB GDDR7 | 16 GB GDDR7 | 24 GB GDDR6X | 16 GB GDDR6X |
| Memory bus | 512-bit | 256-bit | 384-BIT | 256-bit |
| Memory bandwidth | 1792 GB/s | 896 GB/s | 1008 GB/s | 736 GB/s |
| TGP | 600 W | 400 W | 450 W | 320 W |
| Row 7 – Cell 0 | * rumoured specs | Row 7 – Cell 2 | Row 7 – Cell 3 | Row 7 – Cell 4 |
So yes, that RTX 5080. The rumours are not kind, and neither has been the reaction to its own specs. With a reported miserly 5% increase in CUDA core count over the RTX 4080 Super this new card isn’t looking particularly exciting. With the same 256-bit bus and 16 GB of GDDR7 VRAM there’s very little that’s getting my techie heart pounding about that make up.
It looks like it’s relying almost exclusively on an increase in memory bandwidth and a ton more power jammed through its silicon to be able to gain the performance lead over the RTX 4080 Super it needs to deliver a really tantalising upgrade prospect. I mean, the scuttlebutt suggests that it ought to beat the gaming performance of the RTX 4090 by a touch—though given these specs I’m struggling to see it—but it’s still going to be a tough sell given the gap between the RTX 5080 and RTX 5090 is wide enough you could toss an asteroid through.
As ever, we’ve got to take these specs with a dose of the salty stuff until we get official confirmation from Nvidia itself on its new cards, and it plays them classically close to its chest right up to launch. But they feel right. In a way that, y’know, feels kinda wrong. No-one wants the RTX 5080 to look so gimped, with the possible exception of AMD who might be tempted out of high-end GPU retirement if the target is potentially achievable.
Though we still don’t know when the cards themselves will launch. From the people I’ve spoken too, I’ve long had a feeling it would be this year in some form, but then Nvidia has no reason to have to release the RTX 50-series in 2024. It has no competition, and the RTX 40-series are still doing the goods right now. But even if the RTX 5090 and RTX 5080 are to launch in the next three months there’s a chance that these numbers aren’t final, especially those power figures. We saw a whole host of TGP speculation ahead of the final 450 W reveal of the RTX 4090, so there could still be some movement even if these leaked specs are correct as of right now.