While they are possible (and adding features not clearly apparent by the specs) I really don't agree with those rumors. They don't make much sense on their face, but architectures can drastically change in ways not apparent until we see testing/reviews./white papers/etc.
While I am in-fact pulling this out of my speculative ass, I think 4096/8192 appears much more plausible.
Shrinking N33 and pumping clocks (up to ~3264/24000 at stock) would appear a reasonable and cost-efficient thing to do, could compete with products below AD104, perhaps at a size similar to AD107 (which N33 already beats but at a larger size, granted on an older [then cheaper, but 4nm is now also cheap(er)] process). Shrinking from 6nm->4nm could/should increase density by ~40% (44% by my linear math, but I don't know exactly how much more space the lesser-scaling parts like cache would take up, nor do I know how much space arch changes/increasing clocks would effect it). You're really getting into the weeds when you think about the fact 7nm SRAM cell size is .027µ (no idea on 6nm) and 4nm is ~.0199mm2µ (same size as n3b, n3e is WORSE), while decap (for higher clocks) on rv770->rv790 was ~26mm2 (way back on 55nm).
The processes, clocks, and available ram chips make this very possible, and it's placement makes sense (cheapest option that will get you PS5 performance), and I'd be very surprised if they couldn't make it work with 16GB for <$300.
I personally stopped thinking GDDR7 was plausible a while back as I don't *think* they will be available yet by the time this needs to launch. I *could* be wrong. Samsung has been very quiet about a release date (the only one we have is Micron at the EOY, I assume for Blackwell). It could happen, but I still think 24gbps GDDR6 is more likely due to the fact literally no other products have used it (it should be much cheaper) and it fits rather well with a small update to their current parts. Also, if you do the math, 24gbps is actually a better fit. The ram bandwidth would be a wash (or perhaps overclock slightly better), but unless AMD doubled up their L3 cache it would limit bandwidth more quickly (16MB of L3 constitues about 8.5% of bw perf if such a part is 2720/20000). It just sounds more expensive for very little gain...maybe ~5% and that's IF it could hit ~3700mhz, while losing 4GB of buffer. Not a trade I would make.
Likewise, doubling that up is very much akin to what AMD has done in the past. IMHO the actual shader count is likely to be close to 7800xt as that part is clocked at 2425mhz for a reason (part of which may be to make a similar part with higher clocks look like a generational leap), likely to look like a faster PC alternative to the PS5 pro for similarish (and eventually within it's lifetime) less money, which is something the PC gaming market really needs to do if it wants to survive.
8192*3200/7680*2425 = +40% (not counting whatever architectural enhancements may occur).
These just make the most sense to me. We know that 4nm(P?) can yield well at 3200mhz (base clock of Zen4c, typical operational speed of M2 is 3200-3300mhz), we also have seen it scale to 3500mhz (M2) and 3600-3700mhz (M2 Max/Ultra, Zen4c max boost). To me, it appears like the process is most comfortable around 3200-3500mhz mark, which fits perfectly with 24Gbps GDDR6 (including overclocking to ~25.6gbps) and similar L3 cache size to old parts.
This is bang-on to what TSMC claimed about n4p (6% greater density, 11% greater performance). Most (good yield) 5nm GPUs can hit 2900mhz at 1.08v (or lower). I don't see how ~3200mhz @ 1.2v would be a problem (with conceivable headroom to ~3500mhz), nor how overclocking to the above would be out of the question. If you do the math, (even with overclocking) it would appear to be able fit within 375w (or parts with 1x8-pin on the low-end and 2x8-pin on the high-end). Some of you may say "But that's the power draw of a 4080 Super for conceivably slightly less performance". That's very true, but it could also conceivably be the size of AD104, hence a similar MSRP and beat it's actual competition (4070 super), and conceivably tie (and/or beat in raster) 4070 Ti Super. Essentially this part would make a lot more sense than discounting 7900xt into the ground, likely past the point they would ever want to if they still wanted to make any margains. Likewise, I believe a 16GB (stronger than) PS5-like card would go down very well if under $300 (probably less than a discounted PS5), which just might not be possible for navi 33 due to it's size. It might very-well be possible due to 4nm lowering in cost (and more wafer allocation being avaiable) recently (which it certainly has due to Apple/nVIDIA moving to 3nm) as proven by Sony/AMD moving to 4nm, which they wouldn't do unless it was cost-effective.
I also agree that displacing 7900xtx is a tough thing to believe, which is why I think cancelling N4(c?), the chiplet 9 SED (17280sp?) design, was such a difficult decision on AMD's part, as one can imagine how that could be competative. They could always refresh the current chips with 24gbps memory, and we know they are capable of greater than 3000mhz (which is where a refresh/faster memory begins to make sense for N31), but they indeed would consume a ton of power and probably still not catch up to the 4090 (which nVIDIA themselves could then [and may still] refresh). I suppose if they want to make an overtly compelling 4080 Super competition they could, but to what end when Blackwell isn't too far away and the lowest-end "enthusiast" chip in that lineup could very well match a 7900xtx in raster (my best guess at this point is 192-bit/18GB/9216sp @ 3375/32000). AMD might have a slight advantage over-all conceivably in raster at the same price if you factor in overclocking, but that's where the arguement of power consumption(DLSS/RT) really could become a deciding factor. There's always the possibility of a 4nm respin, but I'm not holding my breath. Would love to be wrong as more options and competition is good for everyone.
Point is, the demarkation line (in my mind) is the PS5 and the PS5pro. To put it in more-simple terms, a 7600xt (which is old process tech and still over the golden $300) or overclocked 7800xt (OT: which some people won't do or consider when looking at review charts...even though you'll likely get close to a 20% gain vs a dirt-cheap MSRP 2425mhz stock model and does become a compelling option versus a 4070 Super, especially given the extra ram. /OT), so really they need a cheaper 7600xt and 7900xt (the latter of which is the gateway to high-end PC gaming). They need stock competition for the 4060 Ti 16GB (preferably under the magic $300 mark, which 7600xt is not and nVIDIA would likely never challenge UNLESS they release a point 4060 Ti Super with 4608sp and relegate the old part lower to compete; both hypothetical super and conceivable '8600xt' would probably OC to a similar level in raster) and 4070 Ti Super (preferably for the price of a 4070 Super or less; 7800xt can't hit that perf level and 7900xt costs too much to make). Those would be that thing. They might use a bit more power or run warmer (albeit most AIB coolers are plenty strong), but would make financial sense (wrt die size/cost) versus a chip one slot down in nVIDIA's lineup, which is typically where AMD likes to play. They've always done that: they price themselves against their weakest aspect versus nVIDIA. In this case it has been RT (by ~20%), which again I find amusing because not only do AMD's chips have stronger raster (especially after adding the delta from overclocking), nVIDIA outdates their last gen wrt RT with every new gen (while raster/RAM springs eternal for actual playability at a given resolution). Seriously, look at the playability of a 2000/3000 series in RT with similar raster performance (that's right, the old cards are largely unplayable)...but that's the market. Win for value if you don't constantly update for RT or to get back the missing ram nVIDIA took from you and that you eventually needed.
But hey, let's see how it pans out. AMD has disappointed before, but they also usually have at least one part that's a diamond wrt $/perf (especially if you overclock).
