Watch out Nvidia – new AMD RDNA 3 GPU leak promises big things

AMD’s next-gen graphics cards are rumored to be somewhat different than previously expected spec-wise, going by a new leak – and while there’s a bit of disappointment about some assertions made in this fresh spillage, there’s really nothing to worry about. Indeed, it’s likely Nvidia that’ll be worrying over some of the info floated here…

Before we get into why that might be the case, let’s examine the leak itself, which comes from Angstronomics (via VideoCardz (opens in new tab)), a publication which covers the semiconductor industry that we’ve heard from before. (Even if it isn’t one of the more regular sources that divulge info on Twitter, we regard Angstronomics as reliable enough – plus a major YouTube leaker more or less backs up these seriously in-depth details, and we’ll come back to that).

The key points to this speculation include that AMD is concentrating on “area, area, area”, meaning smaller chips that are still powerful enough to hit Team Red’s performance targets for RDNA 3, and also a focus on performance per watt (efficiency). Of course the company has already boasted about hitting a 50% increase over RDNA 2 with the latter (the same jump RDNA 2 made over the original RDNA architecture).

Apply your own seasoning liberally with all this, but Angstronomics tells us that Navi 31 – the flagship model for RX 7000 graphics cards – will be as rumored with one Graphics Chiplet Die (GCD) plus six Memory Chiplet Dies (MCDs), with 12,288 cores (known as ALUs). That main GCD will apparently be 308mm² in size, and the other big change from previous speculation mentioned is the Infinity Cache size which is apparently 96MB for the top dog model (a cut-down Navi 31, presumably the non-XT version, will use 80MB).

The rumor mill has mentioned much bigger sizes than this, such as 192MB; and indeed the existing Navi 21 employs 128MB of Infinity Cache, so it’d actually be smaller than that. Angstronomics reckons that AMD experimented with twice the amount of stacked cache for Navi 31, but that this didn’t work out to provide enough additional performance to be worth doing.

Hence some of the slight disappointment around these rumored specs, and the second point which has caused a little concern comes with the purported configuration for Navi 32, which Angstronomics asserts will run with 7,680 cores rather than 8,192 as previously believed. Infinity Cache is apparently set to be pitched at up to 64MB, and maybe 128MB for a 3D-stacked model (but again, it seems like AMD may have mulled this idea, and then abandoned it due to it not making sense cost-wise for the extra performance gained).

As for Navi 33, the mid-range offering, this will be a lot smaller if Angstronomics is on the money, with the chip previously rumored to perhaps be up towards 400mm² in size, but actually weighing in at more like 200mm². It’ll supposedly run with 32MB of Infinity Cache.

Regarding the cooling system for the flagship RDNA 3 graphics card, AMD is thought to be keeping a similar triple-fan system to the existing flagship, but it’ll be a bit taller. The GPU will use a pair of 8-pin PCIe power connectors, according to the leaker.

Analysis: Worried by all this? Well, maybe it’s Nvidia that should be worrying…

As mentioned, some folks have been worrying over elements of this purported spec, and some of the bits which seem to have been cut down slightly – or a lot in the case of Infinity Cache – compared to previous chatter from the grapevine. However, there’s no need to worry really, and indeed, it’s potentially Nvidia that should be doing the fretting, and here’s why.

Interestingly, Moore’s Law is Dead (opens in new tab) (MLID), a prominent YouTube leaker, has swiftly chimed in on this one, and after having spoken to multiple sources, he pretty much backs up everything that Angstronomics puts forward here. That said, there’s one key difference, namely that for the Navi 32 GPU, MLID is still hearing it will be 8,192 cores, rather than the 7,680 cores mentioned in this new leak. Of course, MLID still admits his sources could be wrong, so it could go either way yet (and 7,680 cores is certainly feasible).

MLID’s sources do broadly back up the accuracy of the specs and smaller die sizes mentioned here, and lesser loadouts of Infinity Cache – but here’s the key thing on this front. MLID notes that the performance estimates he’s hearing haven’t really changed, and that he still expects AMD to maintain the rumored performance leaps we’ve heard about in the past – enough to compete with Nvidia’s RTX 4000 range – even if specs have been tweaked along the above lines.

Even if performance does end up a touch weaker than Nvidia, there shouldn’t be much difference MLID theorizes. And we have to remember that given the smaller die sizes mentioned, it’ll be cheaper for AMD to make these RDNA 3 GPUs – and therefore price them to attack Nvidia RTX 4000 models. That could be the most exciting thing about next-gen cards, if AMD gets aggressive with pricing, Nvidia could be in trouble – particularly seeing as its more power-hungry GPUs (at least going by the rumor mill) could mean other complications like a PSU upgrade.

Not to mention that Nvidia is supposedly having a tough time with deciding on the next-gen Lovelace launch schedule, which could allow AMD to get in there first with RDNA 3 graphics cards, availing itself of a head-start for sales on top of all these other factors. We shouldn’t get carried away with any of this rumor peddling, of course, but this latest spillage must be worrying for Nvidia on several fronts, we’d imagine.

Oh, and as for Intel – Angstronomics mentions Navi 33 outdoing the flagship Arc Alchemist GPU while being more power-efficient, and costing less than half as much, which would obviously be a bit of a torpedo to sink Team Blue’s desktop GPUs (which are already struggling badly even as it is, without next-gen rival products being around).