Monday, July 21st 2025
AMD's Upcoming UDNA / RDNA 5 GPU Could Feature 96 CUs and 384-bit Memory Bus
According to one of the most reliable AMD leakers, Kepler_L2, AMD's upcoming UDNA (or RDNA 5) GPU generation will reintroduce higher-end GPU configurations with up to 96 Compute Units (CUs) in the top-end Navi 5X SKU, paired with a 384-bit bus for memory. We still don't know what type of memory AMD will ultimately use, but an early assumption could be that GDDR7 is on the table. When it comes to the middle of the stack, AMD plans a GPU with 64 CUs and a 256-bit memory bus, along with various variations of CUs and memory capacities around that. For the entry-level models, AMD could deliver a 32 CU configuration paired with a 128-bit memory bus. So far, memory capacities are unknown and may be subject to change as AMD finalizes its GPU lineup.
After the RDNA 4 generation, which left AMD without a top-end contender, fighting for the middle-end market share, a UDNA / RDNA 5 will be a welcome addition. We are looking forward to seeing what UDNA design is capable of and what microarchitectural changes AMD has designed. Mass production of these GPUs is expected in Q2 2026, so availability is anticipated in the second half of 2026. A clearer picture of the exact memory capacity and type will emerge as we approach launch.
Source:
Kepler_L2
After the RDNA 4 generation, which left AMD without a top-end contender, fighting for the middle-end market share, a UDNA / RDNA 5 will be a welcome addition. We are looking forward to seeing what UDNA design is capable of and what microarchitectural changes AMD has designed. Mass production of these GPUs is expected in Q2 2026, so availability is anticipated in the second half of 2026. A clearer picture of the exact memory capacity and type will emerge as we approach launch.

201 Comments on AMD's Upcoming UDNA / RDNA 5 GPU Could Feature 96 CUs and 384-bit Memory Bus
Btw, Adrenalin 25.8.1 includes support for Polaris and Vega, meanwhile Nvidia Pascal is losing support in October.Engineering is expensive, so it makes more sense to spend engineering on a product that moves more volume, rather than chase the halo tier where there are very few buyers.
www.amd.com/en/resources/support-articles/release-notes/RN-RAD-WIN-25-8-1-POLARIS-VEGA.html
I also feel they should ignore both xx90 cards, they not normal cards, have obscene power requirements and price points.
I do think AMD made a step forward in their last releases, pure rasterisation isnt king anymore, they need to keep making strides on FSR, and use the xx80 card as target for their top end card. Dont worry about the top 1% of buyers.
No one is dropping $2500 on a card to be in second place.
Better for AMD to compete at the xx80 tier and below unless they know they can beat NV on both hardware and software, although we'd all love to see the battle of the titans to return.
The 80 series has drifted upwards over the years. From $500/600 with the 1080, to $700 for the 2080; 3080 was weird bc of reasons, but it realistically was $800-$1000, then 4080 $1200-then-$1000.
5080 is freaking $1200+, and at times until recently was priced closer to the MSRP of a 4090 ($1600).
AMD can compete ABOVE the 80 tier because their price for 80-tier performance may be what some of us call 'rational', or much lower, while a higher-perf card can compete with 80-tier PRICING (better perf).
While you're right that nobody wants the second-best $2500 card, AMD was weary of making $1200 in the post-4090 era as they knew nvidia would drop prices to $1000 on the 80 series (which they did).
And feared nobody would buy a $1200 4090 competitor because most wanting that performance and willing to spend the money already had. Looking back with 20/20 it *may* have worked, but who knows.
In the upcoming instance, many have not bought a 4090-level card bc they were $1600+, and the price did not go down. Like-wise, 5080 did not replace it in performance. 5090 was $2000+.
So we very much have an instance where nVIDIA truly does want stagnation, and does not 4090-level performance to be cheap.
While the 6080 may *finally* bring 4090-level performance to the 80 series, it's very likely nVIDIA will want to keep any price erosion from happening; meaning at least $1200, maybe even more,
The same will almost-certainly be true of wanting to keep a premium on the 5090-level, and whom knows about what could be higher. Would they dare charge $2500 or even more for a 6090? IDK.
When you look at N48, for example, or even N44, you can see AMD was able to bring ~200/400mm chips with 16GB down to ~$300/600. Yes, current prices are ~350/700, but you get my point.
MLID has spoken about how AMD wants to keep the 192-bit (72CU?) chip below $550, and by extension that would likely put 256-bit chips at around $700-800.
Maybe slightly higher, but surely less than $1000 and whatever nVIDIA's competing sku may be.
-
Why are these prices important? Because it was the successful selling price for not only the 7800xt, but the higher-end 7900xt and 7900xtx...and currently where we find the 9070 xt (as a 7900xt replacement).
Even further, one can extrapolate a 384-bit chip to be $1200-1500; still less than a 4090 but perhaps competatitive with a 5090...which nVIDIA will have to lower in (relative) price from $2000.
IOW, it gives AMD choices to compete with whatever nVIDIA brings to replace the 4090 (6080), as well as one tier above, while allowing price space to compete with those skus for less/similar.
At the top-end who knows what will happen; but the point is those are mostly AI/workstation GPUs anyway, and will likely be sold one way or another. Be it AMD at $1600-2000(+), or nvidia at $2500+.
As I've personally said countless times, I've held on to my ancient 2080ti for ages because I want ~4090 perf but will not pay freaking $2000 for one, and the $1200-1600 5080 did not cut my requirements.
I have no idea what nVIDIA plans to do wrt 4090/5090 performance and pricing for Rubin, but as close to as stagnation would not suprise me.
What also would not suprise me is if AMD attempts to hit the spots 7800xt, 7900xt, 7900xtx, and 9070/xt actually sold well...which is to say ~$500-800. They'll likely shoot higher, but also too be a bargrain.
I know it sounds weird to say AMD might make a 384-bit 5090 competitor versus a 256-bit 4090 replacement in terms of MSRP, but that seriously is the way things are trending.
AMD thrives in that ~500-700/800ish market, and if they do enter a tier above $1000, you can almost guarentee it will be to put the kobash on nVIDIA's inflated prices and stagnation of >4090 perf (for less).
This, quite notably, is the difference between 30-40% margin (AMD) and 70%+ margin (nvidia), especially with higher-end parts.
Not only is it ~$800 vs ~$1000, It's also the difference between a card being $1200 or $1500-1600, or $1500 vs $2000.
If there is competition, that absurd nVIDIA margin really makes itself known, especially at higher overall prices. In many cases, it could seperate two completely different classes of cards, if not a half-step.
I don't want to mix what I've been told vs my own personal conjecture, so I'll try to keep them compartmentalized the best I can.
I was told AMD got cold feet on pursuing a large RDNA4 configuration when it became apparent nVIDIA would cut the price of '80'-class from $1200 to $999 if need-be (and as you know, they did with 4080S).
IOW, they felt comfortable releasing a high-end chip that was more powerful that 4080 (and perhaps 5080), but obviously did not feel comfortable competing with 4090 for multiple reasons.
Fast-forward to now, and not only has the 80-class 5080 kept it's $1200 price tag, but has often risen to 4090 pricing (due to a host of factors, including controlled supply from nVIDIA, etc).
It is my personal opinion AMD now feels more comfortable creating a stack that will go over $1000 as they know nVIDIA will attempt to keep that level of performance as expensive as possible.
This is shown by not only did they not truly replace 4090, but that they raised the price of the proportionally faster 5090 from $1600 to $2000.
I am not saying AMD is not fighting for the lower-end they are known for, indeed it appears they are (perhaps certainly with a 192-bit chip that may replace 9070 xt), as well as perhaps a 128-bit chip (that may also be used in the PS6) for even less, but that that higher-end chips, such as 256-bit, could still be sold for under $1000 but indeed fight an 80-class card priced at $1000, $1200, or even more.
At the same time, if nVIDIA chooses to attempt to raise the price of the 80 series further (Beyond $1200), perhaps closer to the price of 4090 ($1600), which we have indeed seen, AMD would have a better option.
I simply do not agree AMD cannot compete on the higher-end (which may have been true up until this point). The reasoning is that nVIDIA refuses to give better value each generation while AMD may.
When you include factors such as expected increases in RT/PT, as well as FSR evolving to a state at least similar to current FSR4, and while they may or may not attain parity, they almost surely will have lower prices.
For instance, you may see a 256-bit ~$800 AMD GPU competing with $1000-$1200 80 class (~4090) performance, while a 384-bit chip may indeed be priced against that 80-class, but perform similar to a higher-end nVIDIA chip. IOW, while AMD has been shy to fight in higher price brackets, if nVIDIA continues to raise prices or stagnate perf/$, there is an (new) opportunity for AMD there...a cheaper 5090.
Again, I say something like $1200 where nVIDIA may be closer to $1600 (for 5090-class perf), which may indeed be a cut-down 512-bit chip.
Likewise, nVIDIA may shoot for the moon with perf/$ similar to 5090 ($2500?) with their highest-end, where AMD may sell those (full) chips for much less, perhaps competing with nVIDIA's $1600-$2000 part.
TLDR: I get what you're saying about it not making sense to invest into a large/expensive chip that may not sell, but I'm saying nVIDIA's margins are so huge and perf/$ stagnation so bad that AMD may find a way, especially with feature parity becoming more of a reality (with Redstone and whatever software/hardware changes, including Amythst, comes after) they may be able to make such relative values that it makes sense.
Maybe nVIDIA does indeed make a super-fast $2500 card, and it does sell for enthusiasts/AI workstations, and AMD only makes a $1600-2000 card that they cut down to 3/4, but that very well could still sell.
Especially if it's priced against an 80-class card, where-as AMD's class of performance may be much cheaper ($700-1000?). Obviously I don't know the exact prices, but I think AMD could make that value work, especially higher up the stack, now that their features are catching up and nVIDIA has become so incredibly greedy even with 'lower-end' card pricing and/or performance (the latter intended to upsell).
RX 9070 XT is already overpriced because it's usually is about 200$ (+35-40%) more expensive than RX 7800 XT but have almost the same specs (technically very close on paper)
AMD are making more money on RX 9070 XT than on RX 7800 XT. For 800$ it's a rip off or nvidia's everyday life.
You should not look at nvidia's pricing because it has nothing to do with reality at all. Only current nvidia gpu that's right now is somewhat decent is RTX 5070 and for other models you will pay "empty air" tax.
Rough math is something like 100 KGD on a $20000 wafer; $200, ~$300 24GB G7 (to us, obviously it doesn't cost manufacturers that much), and then misc board parts/caps/etc and AMD/partner margins.
It could be less, but I also think many people will buy a cheap(er) 4090+ whatever it is. I think that level of performance (~10-20%+?) is the biggest market 'get' right now, and we need a strong value option there.
I think AMD knows this and will do their best, but also making whatever money they can while still convincing people to take it over nvidia's offering. What that true number is...I don't know.
All I do know is 7900xt sold well at ~$700, and 7900xtx sold well at about $800 (or so). 9070xt is selling well at ~$700 as well (at least for now). It's just a ballpark, and that ballpark is indeed pretty big.
Another reason is that MLID keeps claiming the 192-bit part (18GB and likely still quicker than 9070xt) is aimed at <$550 (like 7800xt). 256/192 = 1/3. 550*1/3 = ~$733.
Higher-end parts tend to have a (slightly) higher margin, partners take different margins (and make price-inflated options, especially on popular SKUs)...
...so I'm just rounding. :P
It's kind of like asking what *any* card is supposed to cost these days. Some are somewhat-close to MSRP, others are almost 2x! So it's hard to really pin down a number when MSRP is becoming truly fake.
Point is, though, $700/800/900/1000....You know nVIDIA is going to try to charge (at least) $1200 again, most likely. Maybe they drop it to a cool G, but all the more reason those lower-number AMD estimates.
I just can't see them selling thr 192-bit part for <$550 (as MLID has stated), and then asking 100% more for 1/3 more performance. Maybe ~50%.
Again, just a guess. Just trying to make a point that you can almost guarentee nVIDIA will release a 4090(+) replacement for $1000 or more, while we *almost* have proof AMD will charge much less.
Maybe something weird will happen and the 192-bit part will cost similar to a PS6 with the 128-bit part (~$500?) while the 256-bit part will cost similar to the next xbox with the 192-bit part (~$700?)
Just a weird non-sequitor, but sometimes things like that make sense. That way console people are happy they have a decent GPU, and GPU users feel they got more for the same price w/o paying for the box.
RX 7900 XTX 384bit UDNA Version - 1000$
RX 7900 XT 320bit UDNA Version - 800$
RX 7800 XT 256bit UDNA Version - 600$
Yes UDNA 384 bit version will not be faster than RTX 6090 but if it's priced correctly then at the end it still will be a better product. For example 35% performance boost for 160% price increase or something similar to that it's just loughtable.192bit version should cost no more than 500$ (in ideal world 480$) because it's closer to **60 class gpu it's cut down small die size. MSRP for RX 7800 XT was also 500$ bigger die and 256bit memory.
RDNA4 @384bit shold be around 40% faster than RX 9070 XT because RX 9070 XT has higher tdp than RX 7800 XT.
So around ~40% boost from 384bit larger die + UDNA 2nm improvements probabbly around another ~40% performance improvement over RDNA4 so in total ~80% performance improvement over a RX 9070 XT. But who knows how much faster UDNA arhitecture + new 2nm node will be it's just a guess..... It also can easily be 70% or 100% faster depending on how good the new architecture is.
Close to RTX 5090 performance after 2 years for 55% less money is a good improvement but AMD may ruin it if it's costs for example 1200-1300$ not 1000$.
384bit is the sweet spot at the higher end 512bit is very expensive, power hungry, giant die size etc.....
IOW: AMD ~4090+ - ~$800
nVIDIA ~4090(+?) - ~$1000
AMD ~5090 - ~$1200
nVIDIA ~5090 ~$1600 (old 4090 pricing)
AMD 512-bit part (192CUs?)/6090 - $1600-2000?
nVIDIA 512-bit prt (192CUs?)/6090 - Damned If I know.....$2500?
It's always possible there could be a cut-down version as well below a full chip that could be something like 88/256-bit (like 7900xt, perhaps with slower ram?) and it too be similar to 4090.
Essentially I think AMD wants to ~4090/4090+ performance down to the prices 7900xtx and 7900xt sold well, or essentially fight the next-gen ~6070/ti with a better comparative chip than 9070xt at similar price.
I'd love for a 384-bit/36GB chip to be $1000, but I think that's asking an awful lot. I hope you're right, and if that is the case perhaps they may find myself as a customer!
As you alluded, perhaps there is room for a 320-bit salvage chip at that price (that would still toast a 6080/4090?) I certainly wouldn't rule it out.
I'm just happy to finally see performance I've wanted for quite some time, that neither nvidia wanted to be in Blackwell or AMD w/ RDNA4, perhaps come down to more 'performance'-tier pricing.
I don't think anyone right now that currently be buying a 9070xt/5070ti would be displeased with a more advanced featureset but similar raster/buffer to 4090 for a similar price.
I think they will be if nVIDIA makes 4080 V5 that finally gets them over that hill and it still costs them $1200 (or more) years later (Even with better dlss/FG).
I think that's the bigger concern than if AMD charges more than $1000 for what would clearly be a superior chip...but perhaps much less people would buy regardless.
Butttt....if they do...It would still be funny to have it priced against said 4080 Mark V and beat the ever living snot out of it (say closer to a 5090 than 4090, even if a cut-down chip [or not] at $1000-1200).,
Whatever happens, and we have awhile to speculate, I'm just glad to see us get over the 4090 bump at <$1600, perhaps even (quite a bit?) sub $1000... That's my main point.
If AMD can do it cheaper (and catch up with FSR/FG/etc), all the better.
FWIW, I think a big part of the argument that may be lost in this conversation is that N48, and likely RDNA5/UDNA, are much bigger chips than something like N32 (for features and decap/buffer for clocks).
It's not just the bus size...that memory controller/core clock capability also appears slightly weaker as well.
I very much think the next gen, at least for AMD, will be able to clock over 4ghz on desktop. nVIDIA may very-well use more (or less) units in certain configs at more/less-efficient clocks.
But AMD appears very firmly back in the high-clock methodology for desktop. This may very-well be different on console/APU. I don't know. Even if they use the same chips, lower voltage/clocks would save a disproportionate amount of power, but ofc that helps make them cheaper in them first place (less silicon/less power-efficieient, but often cheaper to make), so using the same chips would kind of be a waste if that were the case versus using low-leak (smaller) xtors. No more of a waste than the low clocks in something like the same chip in a laptop though, so I guess in that respect it would be nothing new.
The reason I mention the clocks is because it's been shown in Linux the RDNA4 clock domain very-much scales above 3900mhz...it's pretty much ready for N3E.
Give it N3P/X and 4ghz is certainly possible...and would better fit with the capabilties of 36gbps GDDR7.
I don't think they will use N2 for GPUs (at least not any time soon) as wafer cost is 50% more and the chip sizes bigger and therefore yields are much worse....They would be pretty expensive (unlike CPU chiplets).
Now, if AMD were using chiplets for RDNA5...maybe...but I've yet to hear confirmation they trended back in that direction (which isn't to say they did or didn't). Magnus would *imply* mostly monolithic logic.
To close the feature set gap, make sure that all extensions are fully functional, reduce power and die area, ship the product with less memory and reduce the price while expecting the same performance level... I just won't say it's impossible because it's the tech world, and everything is possible with enough engineering and investment here. But that it will be an exceptionally tall order, it will.
And then there's Bolt Graphics and their outlandish claims. I write that off even more than this scenario, but if by some miracle they deliver what's promised? Whole market is going to be upended. Interesting times ahead...
High end consumers will always chase the 90 series equivalents for the best of the best.
The bulk of mainstream gamers and even to an extent enthusiast class consumers look for an solid mid range card or an upper tier card but not flagship.
Regular gamers would look at a 60 class cards and the buy them,
Enthusiast/harder core gamers look at 70 class cards, some spill into 80 class
Deep wallet people look at 90 class.
Only need to look at (when supplies are plentiful) how many 60/70 class cards are made and sold compared to 90 class.
Go into a shop where they sell a lot of tv's, the biggest most expensive units don't sell anywhere near as much as the solid performance better value mid rangers.
If AMD drop a close to 4090 class card with the next generation for 9070XT money they'll kick a huge goal, it's not the fastest, it's not even previous gen flagship, what it is, is a substantial jump forward in performance compared to what that money buys today, so it's a compelling value upgrade.
I'm admitted an AMD fan, but I wouldn't buy a 5090 if I had deep pockets, unless it's MSRP was 1500 USD and it was priced at that (or in current market, a bit less than 5080's are selling for now)
And this is with an old generation CPU which means very high CPU bottleneck, too, for the RTX 5090.
RTX 5090 is 52% faster @4k but costs 130% more. To me RTX 5090 definitely is worse product than RTX 5080 where you will pay a lot just for empty air. And i will repeat RTX 5080 is already a weak product.
Where is RTX 5080 Ti ? Half less expensive and only slightly slower.
GTX 1080 Ti 8 years later.
ARF angry for what ? Do you really like to overpay just for dumb naming? Nvidia are deliberately not releasing RTX 5080 Ti class gpus anymore to make more money.