NVIDIA's RTX 5090 could be with us late this year - or that's the rumor, anyway (though some disagree) - and we've just heard a bit more about how one YouTube leaker feels this graphics card might pan out.
In a new 'Broken Silicon' episode (246), Moore's Law is Dead (MLID) touches on the topic of the RTX 5090, and retreads some of the rumors aired in the past about the Blackwell flagship, while adding some fresh theories into the mix.
MLID replies to a question about whether the RTX 5090 (built on the GB102 chip) will be cut down for CUDA cores to a greater extent than the RTX 4090 (AD102), or about the same, or whether it'll be closer to the full loadout of cores than the current Lovelace flagship GPU.
The YouTube leaker reiterates that the expected performance increase for the RTX 5090 is that it'll be 60% faster than the RTX 4090 - or maybe even 70%, best-case scenario (and certainly the card will boast a 50% generational uplift).
And to achieve this, it's likely that NVIDIA won't cut down the CUDA cores from the full loadout as much as Team Green did with the RTX 4090.
At the same time, MLID clarifies that the RTX 5090 won't be that close to the full complement of cores either, so it will be more dropped down than the RTX 3090. In short, it'll be in-between these two previous flagships somewhere.
NVIDIA certainly won't want to get too close to using all the cores, as the full die GB102 chips will be needed for AI graphics cards. Products that'll be far more profitable than the RTX 5090, of course, by a large margin.
Even though the RTX 5090 is hardly going to be cheap, and MLID touches on this in the above video, too. The YouTuber predicts that pricing for the RTX 5090 is likely to weigh in at the $2,000 to $2,500 mark in the US.
Read more: NVIDIA's RTX 5000 GPUs on track for Q4 2024 launch but don't expect a Lovelace performance leap
Stock crash?
There's potentially worse news, as what could push up pricing even more is stock being thin on the ground - a possibility MLID mentions, although just as a potential scenario, rather than any firm rumor.
When discussing that NVIDIA will look to cut down GB102 for the RTX 5090 as much as it can get away with, MLID poses an alternative idea - that Team Green may go for a more fully fledged core count (better than the RTX 4090, in terms of the percentage of the full die used, as already noted), but just not make that many flagship Blackwell graphics cards.
In other words, this would be another way for NVIDIA to 'afford' the cost of the RTX 5090 in terms of its production capacity, without taking too much away from the AI arena where the real profits lie.
Could that happen? It's not something you can rule out, of course, and it'd be a nasty approach for enthusiast PC gamers to deal with. The RTX 5090 would sell out in a flash, given this kind of scenario, with the likes of price gouging meaning that the mentioned $2,000+ price tag would be inflated well above that in the scramble for a Blackwell flagship.
We should stress that MLID isn't saying this will happen, or that he's heard it might - just that it would be one possible tactic for NVIDIA: Launching a GPU to dominate the benchmark charts, more than anything else (as opposed to racking up sales).
Obviously we're too far out from launch to be talking pricing and stock levels seriously, anyway, but you can't really argue that this is one possible avenue Team Green could take with the RTX 5090. We'd do well to remember that AI is the key area for NVIDIA without a doubt now, to the point that there's been theorizing about whether RTX 5000 might even be the end of the line for the GeForce gaming family of graphics cards.
Read more: NVIDIA's RTX 3060 could be staying around as a budget option to tackle AMD's rival low-end GPUs