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
As for flawed dies, they will need plenty to create a new model. If they don't have plenty they might create again an OEM product that will go to 1-2 OEMs or be sold to China for example.
This way two full fledged 450mm2 96CU RDNA5 dies can potentially act like one 900mm2 die, but without the extra sync overhead of the chiplet approach.
It would be a sort of hybrid approach somewhere between chiplet and crossfire.
The argument that favors nVidia is the feature set, people say that DLSS is superior to FSR4 even though they can't discern it while actually playing, when you point this out to them, they say yeah they're very close in quality but the game support is far greater for nVidia.
So AMD can't win without parity (for game support), or at least being close enough which at this point it isn't.
There are still people that consider nVidia better at ray tracing even though the real advantage it has is in path tracing.
AMD is forced to make a flagship that can hang in there with nVidia's one, if it does then it proves that nVidia can't deliver, all that money, all those resources and it can't decisively win, if FSR gets closer to DLSS it won't matter if there's a difference if it's only revealed by VMAF tests.
AMD needs to concentrate on game support and getting all of these "productivity" features straightened out and deliver a cohesive package, yes no CUDA obviously, and nVidia will still lead in rendering but if it's just that and the rest is on par then the value of AMD GPUs will be undeniable.
And they've made mistakes this generation, small ones but which got noticed. Like the memory temps, it just makes the products feel second rate, even if they run within specs, and when Samsung fixed the temps they "unfixed" the latencies and performance in synthetics was marginally affected.
It just looks amateurish, without this issue they could have had a selling point, the RDNA4 cards apparently can have great GPU temps with smaller coolers than comparable nVidia cards.
And AMD could have said, hey look our cards have similar thermal and acoustic performance, but they're more compact and lighter and have more reliable power connectors.
But how can you say your product is more reliable when the temp delta between the core and memory is ~30°C???Sure they do. The ones that got 5090s for $4000+, when they upgrade to 6090, 7090 or 8090 they would prefer to pay $3000. :laugh:
Intel, like AMD, relies on TSMC and has even lower priority, Nvidia isn't interested because they have the premium segments in a chokehold, and the Chinese products from the likes of Moore Threads and Glenfly aren't competitive. While this could theoretically be solved by restarting production on GPUs that retain advanced feature sets but are built with older nodes (such as RDNA 2 or Ampere), there is zero business incentive to do so, as the nodes can be used to supply semicustom demand effortlessly (i.e. game consoles and automotive sectors) and it'd only serve to deflate prices on their latest products and stabilize the market, which is something shareholders are ultimately not interested in.
So unless they can make a GPU that pulls all the stops and ticks every box, and can be sold for an equally high price, they aren't gonna make one. To the hyperconscious gamer with a small wallet, an RTX 5090 might sound like an enormous investment... Nvidia is selling these at almost charity prices to gamers compared to what they could fetch by making AI accelerators for the enterprise segment.
The answer to everything is "wait for the reviews"
This means the new architecture will have much beefier compute units, rather than thin ones but with high parallelism.
The UDNA x1 part will probably have 100 billion transistors, and hopefully will be manufactured using the newest TSMC N2P lithography node.
Don't let the people on tech forums cloud your judgement we all a very small part of the market.
Honestly, this makes no sense - it makes sense if the average joe is a die-hard football team fan, but IT people behaving illogically is a mystery. Maybe they are not intelligent, after all?
Then again, the 5050 is rather dried up in that department currently, so maybe there's merit to some of the most popular cards also being the absolute cheapest.
Again I'm not saying there's a functional problem, the cards run fine, it's just that perception wise they might seem lacking for some people, like the case with the feature set, the whole driver debate etc.
AMD has to fight on all fronts with UDNA, it has to dominate against their previous products (RDNA4), dominate against previous nVidia cards (Blackwell) and severely disrupt the used market (make people prefer buying new UDNA cards instead of used Blackwell and used RDNA4) and (at least) match the RTX 6000 series for the whole product stack.
Otherwise they'll just pick up scraps like jackals around a lion kill.Old dusty ones. That are matched by 64 of the 9070 XT. How would a 96 CU RDNA4 card perform?
Then how would a 96 CU UDNA card perform?
AMD will never admit that they've made a mistake by not making a bigger die this generation, but it's so obvious that we don't need them to admit.
But perhaps this generation was a sacrificial one, one where they made great progress but also some mistakes, the important thing is that they got their foot in the door and if they learn from these mistakes then UDNA will be Radeon's Ryzen moment. It's not impossible, not even improbable. The proof is that Blackwell is a stop-gap solution while nVidia waits for 2nm. While RDNA4 is clear progress compared to the previous gen. It's easier to miss the progress because the product stack stops at the midrange.
Anyways, and aside from that, I was misremembering the CU count of Navi 48 being 48 and not the correct 64, so yes unlikely to be faster than a 5090 unless its a hair.
But even IT people like to pick teams, it makes no sense to me.I think AMD getting a foothold with RDNA4 is much more important than chasing after the high end, especially when the high end buyers will open their wallets for an Nvidia card regardless, just see the comments from people here saying they'll buy a Geforce card no matter what.The same people who always derail these threads can't see the irony in it.
If the CU's are even more beefed up than RDNA4, as RDNA4 has higher IPC than Blackwell, AMD does have a chance with competing with RTX 6000 series.
Though I'd be more interested in a chiplet design midrange UDNA card, the high end has gotten so expensive it's a niche of a niche of potential sales.
RDNA3 was a really cool proof of concept, RDNA4 was originally supposed to be a chiplet design but everything larger than the N44 got scrapped and replaced by the last minute with the N48, presumably because it was running into the same clockspeed issues RDNA3 did.
Presumably N5x was still being designed as a chiplet based arch so if AMD has had time to smooth out some of the clockspeed issues associated with RDNA3, there is still a chance for RDNA5/UDNA to be chiplet based to deliver the same 800mm2 of silicon NV does, but 400mm2 at a time.
Do you know why nobody tests fill rate or texture rate anymore? Because raster is obsolete. It’s irrelevant.
Node shrink for higher clocks is good, 50% increase in CU's, also good, but if the ipc remains the same then even flagship next gen will fall behind nvidia's current gen flagship.
I think in order to be attractive it needs to be able to trade blows with or be faster than a 5090/6080 but at more than a 5-10% discount.
In my market right now (Australia) the 9070XT trades blows with the 5070Ti but is 20+ % cheaper.
Regardless of the tier they end up competing at, if they're not at least 20% cheaper they're not going to entice many nvidia owners away (or enough to swing the market share back a little)
I will throw this out there though, for the "FSR4 only supported by a single generation of cards" crowd, remember DLSS didn't work on every generation of nvidia card nor does every feature of it run on every RTX card. Next gen AMD cards will have 2 generations of FSR4 cards. Nvidia will have 5 for DLSS.