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
Add your own comment

201 Comments on AMD's Upcoming UDNA / RDNA 5 GPU Could Feature 96 CUs and 384-bit Memory Bus

#76
GodisanAtheist
colossusrageblackI see this maybe matching a 4090, highly doubt it'll match or beat a 5090. At $1000 though, matching a 4090 isn't bad. It's what we expected the 5080 to do, maybe this will light a fire under Nvidia.
-A 96CU 384-bit RDNA4 die would already be comfortably inbetween a 4090 and 5090 in most workloads today.

If RDNA5/UDNA/Whatever comes with an IPC uplift it should be able to match or exceed the 5090.

Big If tho.
Posted on Reply
#77
tpa-pr
GodisanAtheist-A 96CU 384-bit RDNA4 die would already be comfortably inbetween a 4090 and 5090 in most workloads today.

If RDNA5/UDNA/Whatever comes with an IPC uplift it should be able to match or exceed the 5090.

Big If tho.
The thing is (for my purposes), I don't need it to be a 5090 equivalent or top of the charts. I just need it to be a significant uplift over my 7900 XTX. I buy AMD for the price and the software stack (the Linux software stack that is).

I also don't think there's anything wrong with AMD not being at the top of the charts. As long as they provide relatively powerful hardware at a decent price I think that works fine. Besides, even if they did beat the 5090 I don't think it'll cause a drastic shift in marketshare in their favor. The Nvidia mindshare in AI and gaming is just too strong. There are other ways they can chip away at the marketshare but just being "the alternative" to the top dog is fine, as long as they make money.
Posted on Reply
#78
Dr. Dro
GodisanAtheist-A 96CU 384-bit RDNA4 die would already be comfortably inbetween a 4090 and 5090 in most workloads today.

If RDNA5/UDNA/Whatever comes with an IPC uplift it should be able to match or exceed the 5090.

Big If tho.
Positively yuge "if." To achieve the same level of performance of the RTX 5090, this purported GPU would need to be ~100% faster on average compared to the 9070 XT. It's a very tall order, no matter how one would slice it.
Posted on Reply
#79
RaceT3ch
Dr. DroSales on 7900 XTX were historically poor, but they picked up well after DeepSeek's public release. I think they sold almost as much as the 4080 overall. Which is a ton less than the 4090 did, but a decent amount nonetheless.
What about the 7900 XT, 7900 GRE and the 7800 XT?
Posted on Reply
#80
Visible Noise
ThunderslatThat's just being ignorant.
Name a recent game that has been ROP limited before you call me ignorant.

Just one, released in the last three or four years.
Posted on Reply
#81
Dr. Dro
RaceT3chWhat about the 7900 XT, 7900 GRE and the 7800 XT?
Check your DMs :)
Visible NoiseName a recent game that has been ROP limited before you call me ignorant.

Just one, released in the last three or four years.
Unless at extremely high resolutions, fillrates have been generally enough for quite a few generations now. I think this was an area where AMD actually had a generational lead, too.
Posted on Reply
#82
Visible Noise
Dr. DroUnless at extremely high resolutions, fillrates have been generally enough for quite a few generations now. I think this was an area where AMD actually had a generational lead, too.
Yep, and that’s why some defenders are stuck in the past. Their brand was good at something then, so there’s no way it could be irrelevant now.

For funsies check his post history. Nvidia bad, AMD good. So predictable.
Posted on Reply
#83
GodisanAtheist
Dr. DroPositively yuge "if." To achieve the same level of performance of the RTX 5090, this purported GPU would need to be ~100% faster on average compared to the 9070 XT. It's a very tall order, no matter how one would slice it.
-At 4K, based on TPU's relative performance charts in the latest 5060 review, the 5090 is ~80% faster than the 9070xt.

At 1440p, it's 60%.

It's a big gap for sure, but with 50% more resources and an IPC boost it would be competitive, at least at some resolutions.

In a dreamworld AMD would figure out how to stitch two of these dies together and go big with their performance but again that's just dreaming.
Posted on Reply
#84
Dr. Dro
GodisanAtheist-At 4K, based on TPU's relative performance charts in the latest 5060 review, the 5090 is ~80% faster than the 9070xt.

At 1440p, it's 60%.

It's a big gap for sure, but with 50% more resources and an IPC boost it would be competitive, at least at some resolutions.

In a dreamworld AMD would figure out how to stitch two of these dies together and go big with their performance but again that's just dreaming.
We're currently held back by CPU limitations at the most common resolutions, and in the case of the 5090, that includes 4K. But with Intel stagnated and AMD just slowly iterating, future is grim here :(

I have a 3 year old processor that has no real upgrade path. "14th Gen" is just a re-release of the same thing, Arrow Lake is a performance regression, Zen 4 and 5 offer nothing noteworthy for gaming, X3D isn't worth the platform swap...
Posted on Reply
#85
Hecate91
tpa-prThe thing is (for my purposes), I don't need it to be a 5090 equivalent or top of the charts. I just need it to be a significant uplift over my 7900 XTX. I buy AMD for the price and the software stack (the Linux software stack that is).

I also don't think there's anything wrong with AMD not being at the top of the charts. As long as they provide relatively powerful hardware at a decent price I think that works fine. Besides, even if they did beat the 5090 I don't think it'll cause a drastic shift in marketshare in their favor. The Nvidia mindshare in AI and gaming is just too strong. There are other ways they can chip away at the marketshare but just being "the alternative" to the top dog is fine, as long as they make money.
Agreed, a solid midrage to upper midrage is all I need, it's all most people need as AMD has sold a lot of 9070/9070XT cards. AMD drivers have been nice in Linux as well, only experimenting for now as I've always used Windows, however MS has really lost their way with a bloated OS forcing ads and AI, IMO Linux is worth a try.
Topping the charts is great for the money is no object enthusiast niche buyers, but I don't see the point of AMD making a x90 competitor when said mindshare has only gotten stronger. Offering a very competitive midrange option has proven to be successful with RDNA4, and there is more money to be made in the low end to midrange. AMD just needs to keep it up with RDNA5/UDNA.
Visible NoiseFor funsies check his post history. Nvidia bad, AMD good. So predictable.
The irony just keeps going.
Raster is still the base for every game made, and will be relevant until every game is fully path traced.
Posted on Reply
#86
Visible Noise
Hecate91The irony just keeps going.
Raster is still the base for every game made, and will be relevant until every game is fully path traced.
You’re invited to answer the question also, since Thunderslat couldn’t. Name a recent game that is ROP limited. You know what a ROP is right?

We’ll wait.
Posted on Reply
#87
alwayssts
Perhaps I'm being ignorant, but I don't really see why #most consumers need something faster than a 4090 for quite a while, if ever. The main concern is not raster/rt/pt, but the up-scaling model overhead relegating it lower over time.

Look at games; pretty obvious the high-end has been normalized around this spec for a while, and I don't see any reason that wouldn't continue given the slowing/cost of both process tech and unit advancement per generation.
Remember that 5080 does not catch the spec of a 4090; which is to say 60/120fps mins at 1440p/4k or 1440p->4k with DLSS in demanding titles. Hence, this spec is ever-relevant; certainly within ~375w.

The only major areas of improvement over 9070xt needed are RT/PT limitations (say something like Wukong that would need a ~1/3 improvement over 9070xt + whatever overhead of FSR to scale 1440p->4k [15%?]).
Overall 50% should do the trick to parity 4090 me-thinks; maybe even surpass it given AMD doesn't play the "make sure old cards can't maintain 60/120fps by updating DLSS" game (which is clearly present on 4090 transformer, if slightly).

I know they kinda-sorta did it with RDNA4/FSR4, but we all know how both companies *generally* operate and AMD would likely one-and-done this spec without forcing future unnecessary upgrades (unlike 4090 through new DLSS/features)

Say something like 12288sp @ 4ghz. 1/3 more bandwidth for clocks + whatever extra RT/PT would use. Given RT on 9070xt appears to use ~5(-~10?)% extra bw, and 3nm ~4-4.5ghz, lines up with GDDR7 specs (32gbps?).

Very achievable. Even if PT requires even more bw (which it does), faster GDDR7 is available and could/should allow a decent experience (perhaps within the VRR window for demanding titles?).

It really is about making a 4090 that won't be outdated by generational/updated features for it's absolute capability (like MFG or newer dlss overhead which chips away at older Geforces) and simply making it more affordable/available.
This is both nVIDIA's and AMD's M.O's, respectfully. Bring forth thee finest of wines.

I'm quite okay with that, especially if next-gen consoles are say perhaps about upscaling ~1080p-1440p->4k (performance/balanced, maybe some 'quality'), and a card like this could maintain 1440p->4k with bells/whistles at intended FR.

Another thing to remember is even if you're opposed to owning an AMD card, this will likely push nVIDIA to make a stronger 6080 that can actually compete/surpass 4090 (even if they increase DLSS5+ overhead to eventually outdate both).

I know some of you think me cynical, but that's the likely reality. Win-win, and driving what may more-or-less be for all intents and purposes consumer endgame down from it's past, current, and future inflated prices.
Posted on Reply
#88
Dr. Dro
alwaysstsPerhaps I'm being ignorant, but I don't really see why #most consumers need something faster than a 4090 for quite a while, if ever.
Who needs more than 640 KB of RAM? ;)
alwaysstsThe main concern is not raster/rt/pt, but the up-scaling model overhead relegating it lower over time.

Look at games; pretty obvious the high-end has been normalized around this spec for a while, and I don't see any reason that wouldn't continue given the slowing/cost of both process tech and unit advancement per generation.
Remember that 5080 does not catch the spec of a 4090; which is to say 60/120fps mins at 1440p/4k or 1440p->4k with DLSS in demanding titles. Hence, this spec is ever-relevant; certainly within ~375w.
Mostly because the RTX 5080 is the actual flagship, the 90 is just for clout. It's more more of a "because we can" than an integral part of their lineup, it's mostly that Nvidia retired their Titan brand.
alwaysstsThe only major areas of improvement over 9070xt needed are RT/PT limitations (say something like Wukong that would need a ~1/3 improvement over 9070xt + whatever overhead of FSR to scale 1440p->4k [15%?]).
Overall 50% should do the trick to parity 4090 me-thinks; maybe even surpass it given AMD doesn't play the "make sure old cards can't maintain 60/120fps by updating DLSS" game (which is clearly present on 4090 transformer, if slightly).
Parity 4090 at a point where it'll be half a decade later on a much more advanced node is hardly a monumental achievement, even if at half wattage, that's the only problem. But I adhere to the school of there are no bad products, only bad pricing. You'll sure have fun with one, and that is what ultimately matters.
alwaysstsI know they kinda-sorta did it with RDNA4/FSR4, but we all know how both companies *generally* operate and AMD would likely one-and-done this spec without forcing future unnecessary upgrades (unlike 4090 through new DLSS/features)
The RTX 40 series are fully compatible with DLSS 4, including frame generation. Blackwell additionally supports triple and quadruple rate frame generation, made possible thanks to a hardware-assisted frame metering engine. This isn't present on Ada, however, so it cannot be backported to the 40 series. Game support is currently poor, with few games supported. Otherwise, every GPU as far back is Turing is fully compatible with DLSS's super resolution and ray reconstruction functionality, no buts or ifs.

It is possible for Nvidia to backport DLSS 4's frame generation support to the RTX 30 series, but there is no market incentive to do so. The 40 will officially receive Smooth Motion support soon, already available on a developer release 590 branch driver.
alwaysstsSay something like 12288sp @ 4ghz. 1/3 more bandwidth for clocks + whatever extra RT/PT would use. Given RT on 9070xt appears to use ~5(-~10?)% extra bw, and 3nm ~4-4.5ghz, lines up with GDDR7 specs (32gbps?). Very achievable. Even if PT requires even more bw (which it does), faster GDDR7 is available and could/should allow a decent experience (perhaps within the VRR window for demanding titles?).
Not at all achievable. Assuming that a similar WGP layout is kept, 12288 SP would translate to 192 CUs. This will result in a processor that is roughly triple the size of the 9070 XT, and you'd need a process node that also triples the mTr/mm² rating of N4P to achieve it at roughly the same size. Needless to say, this is beyond the realm of the wildest of dreams with technology we'll have during the 2020s, if not the 2030s, and it certainly won't be an affordable node to develop and sell products on for the foreseeable future.
alwaysstsIt really is about making a 4090 that won't be outdated by generational/updated features for it's absolute capability (like MFG or newer dlss overhead which chips away at older Geforces) and simply making it more affordable/available.
This is both nVIDIA's and AMD's M.O's, respectfully. Bring forth thee finest of wines.
The RTX 4090 has not been neglected and it hasn't been denied any new features whatsoever, and the transformer model upgrade to DLSS (preset K) is neither mandatory, default, or has any major impact on performance, even in earlier generation hardware. DLSS preset K will run on GPUs like the 2060 and 3050 just fine, and while the transformer presets are slower to infer, this is because the model is more complex and the hardware isn't as fast. It's still only about half to one ms per frame slower than the convolutional presets, so it's not really a big deal. The performance reduction from forcing FSR 4 on unsupported hardware seems an order of magnitude higher than that from what we gather from OptiScaler and earlier tests on Linux, and even then I'd argue AMD should have released it to everyone as it looks better and is still faster than native.
Posted on Reply
#89
jackhackett
Amd claimed a reasonable MSRP, but that was only a tiny number of launch day cards, that Amd subsidized, and had to pay rebates to manufacturers. If they had to give rebates to sell Msrp, then the price Amd gave was likly never going to happen, but they still claimed they didn't lie about Msrp, but that launch day price never was seen again. Nvidia is focused on A.I, and dosen't care what happens with their consumer graphics cards. It as if they gave the consumer graphics card division to the least experienced employees, and told them to do the best they can, without ever checking on them again. The Nvidea CEO claimed it was nessarary to raise prices by 2 or 3 times previous cards to help his poor, suffering shareholders. It's easy to imagine CEO Jensen Huang trying to comfort his shareholders who are depressed and crying because they can only afford gold plated sewage pipes in their mansions instead of solid gold.
Posted on Reply
#90
Hyderz
Lets guess what amd is cooking this time?
Will it be cooked just right leaving everybody satisfied?
Will it be overcooked, leaving people with a slight upset stomach?
Will it be undercooked, leaving people unsatisfied and question its value?
Will it be some 3 Star Michelin cooking, that is impressive to look at but it lacks servings leaving people unsatisfied?
Posted on Reply
#91
mikesg
It's the FP4 performance of the latest CDNA architecture that will come to consumer GPUs that's the benefit, if anything.

The MI350X has a 128 multiplier for FP4 performance (from FP32 base). If the 9070XT had that, it would have 6228 FLOPS of FP4, instead of the ~600 FLOPS of INT4 it has now.

So if the 9070XT was UDNA today it would have 10x the AI and upscaling performance. RAM performance limited.
Posted on Reply
#92
cinemaware
blacksea76You sound like you are rich so good luck with nvidia :)
I so wish spending 100-200$ more on a GPU meant I was "rich" :)
That aside from the fact that for the better part of the year 9070XT prices matched 5070Ti and are only now coming down a bit.
Posted on Reply
#93
dyonoctis
igormpThe X100 chips are solely meant for DC, yes, the same happened with the V100, A100, H100, etc etc. Those were also the only HPC-capable chips since they had proper FP64, unlike the other parts.
However, the GB202 is still widely used in the RTX PRO 6000, which is a really great DC part for inference and even small-scale training. It's also being widely used in workstations.
Yhea, but following the OP logic, that would mean that Nvidia stopped making "gaming GPU" since the 8800 GTX, because the Tesla C870 is a 8800 GTX without display output meant to be used in datacenters :D
www.techpowerup.com/gpu-specs/tesla-c870.c1542
www.techpowerup.com/gpu-specs/geforce-8800-gtx.c187
It just feels like some people have a fantasy about a wicked-fast and wicked-efficient gaming-only GPU, while forgetting that even the beloved Maxwell was found in data centers. And it even uses the same dies as the desktop variant.

Posted on Reply
#94
LittleBro
Visible NoiseHave you been living in a cave for the last decade?

Do you know why nobody tests fill rate or texture rate anymore? Because raster is obsolete. It’s irrelevant.
Visible NoiseName a recent game that has been ROP limited before you call me ignorant.

Just one, released in the last three or four years.
Any game that can run better with more ROPs is in fact ROP limited.
ROPs still have a meaning. Just go look at perf. drops with ROPgate affected RTX 5000s.
8 fewer ROPs on RTX 5090 caused 6% performance drop in 4K. On 5070 Ti perf. drop was nearing 10%.
You can't apply DLSS/FSR/XeSS and (M)FG without rasterization, they need output from rasterization in order to work.
But maybe in your world, games will AI-guess game scenes entirely, right? Yep, every frame a entirely different (guessed) scene.
Posted on Reply
#95
Aphid
I'd like to see a 72GB 384-bit card at the high end for <$2K, ideally around a 50% price premium on the 36GB version, which would still be a 50% margin on that added memory for AMD.

This could be done by clam-shell and 12 3GB modules on either side of the card, clocked a bit lower to stay within power and thermal limits.

The modules themselves should not be that expensive. At most $20 per module. So that's $240 in memory added to a card presumably priced around the $800 to $1000 mark, for a final product costing $1200 to $1500.

This would be great for a market that's entirely ignored by nVidia: people who want to do local AI cost-effectively need lots of VRAM, more compute than CPUs can offer, but don't need the crazy power and wattage of the instinct cards. The best card right now is still the 3090, or maybe a second hand mi60. Quantization can cut models in 1/2 to 1/4 size without much performance loss, but today the difference between datacenter nodes (192GB or 288GB x8) and consumer/workstation (32GB x4) is too large to be able to run the open models being created in any reasonable capacity. The tendency to go to MoE (sparsity) over dense means they're getting even bigger than before, trading memory capacity for more throughput at equal quality. The issue is that our hardware is doing the opposite: memory capacity is kept artificially low while compute speed is getting faster.

Nvidia's bridging the gap with their 96GB workstation card... but charging a ludicrous price premium of 6,000 for the privilege of putting a few extra $20 memory chips on a 5090 that is already priced at three times cost.

At 72GB per slot, a 4-slot system can run 288GB, equivalent to a single datacenter card, and an 8-slot system can run 576GB, equivalent to two data center cards. The latter could locally run deepseek-v3 at a fast speed. Something that currently takes 5x the price that this product would unlock.

It's either that or add DDR5 slots to your GPUs so the model can at least be accessed without going through the slow PCI-e bus.
Posted on Reply
#96
Doctor Who
john_128bit = 8GB VRAM for the "low end" costing $300?


AMD FSR 4 Now Available in Any DX11/12 Title via OptiScaler | TechPowerUp
People who usually care about DLSS are those who CAN play with such tools.
8GB is useless regardless of cost.
The absolute minimum should be 12GB for the low end, 16GB for the mainstream & anything above 16GB for the high end.
Visible NoiseHave you been living in a cave for the last decade?

Do you know why nobody tests fill rate or texture rate anymore? Because raster is obsolete. It’s irrelevant.
Don't be silly, Rasterization is still used as a core rendering method in pretty much ALL video games.
Claiming its obsolete, irrelevant is nonsense.

Its about time AMD moves to a Unified design. It was far too costly to come up with different designs for different applications.
Posted on Reply
#97
TumbleGeorge
Visible NoiseYou’re invited to answer the question also, since Thunderslat couldn’t. Name a recent game that is ROP limited. You know what a ROP is right?

We’ll wait.
Id there benchmark in use in reviews that shows how much depends of games especially to raster operations number?
Posted on Reply
#98
Unregistered
Without knowing pricing, i have no interest in it. If it's the same strategy like RDNA4 it will have the same success, none.
#99
GodisanAtheist
Dr. DroNot at all achievable. Assuming that a similar WGP layout is kept, 12288 SP would translate to 192 CUs. This will result in a processor that is roughly triple the size of the 9070 XT, and you'd need a process node that also triples the mTr/mm² rating of N4P to achieve it at roughly the same size. Needless to say, this is beyond the realm of the wildest of dreams with technology we'll have during the 2020s, if not the 2030s, and it certainly won't be an affordable node to develop and sell products on for the foreseeable future.
-Very achievable... Because of semantics and not engineering. AMD's SPs are now capable of the same dual FP32 throughput as Nvidia's have been since Ampere. This technically means each SP is actually 2x SP for the purposes of compute power.

So 12,288 SP in this context is actually 6,144 RDNA 3/4 SPs .

Bringing dual issue throughput on par with NV is a major area where AMD can improve its IPC. It's had dual issue since RDNA3, but the actual hit rate for dual issue ain't great.

Of course throughput for dual issue SPs isn't great even in the best of circumstances, which is part of the reason we didn't see Ampere completely leave Turing in the dust despite official specs suggesting massive gains in SPs.
Posted on Reply
#100
igormp
dyonoctisYhea, but following the OP logic, that would mean that Nvidia stopped making "gaming GPU" since the 8800 GTX, because the Tesla C870 is a 8800 GTX without display output meant to be used in datacenters :D
I personally consider that the beginning of CUDA's history, and the GTX 500 series being the first mark on its usage in scale for AI (due to the alexnet paper).
From Pascal onwards then Nvidia started to add more and more features that were compute-orientend, and not really solely gaming-orientend. And the rest is history.
dyonoctisIt just feels like some people have a fantasy about a wicked-fast and wicked-efficient gaming-only GPU, while forgetting that even the beloved Maxwell was found in data centers. And it even uses the same dies as the desktop variant.
For sure gamers would love that, too bad they don't pay enough to have it :p
Nonetheless, I kinda agree with the OP you were replying at first, current non-x100 dies are planned first and foremost for their compute performance, and then gaming is kinda of a 2nd thought.
The x100 chips are solely meant for enterprise and raw compute, no graphics whatsoever (so not even pro graphics like in proviz). But that doesn't exclude what I said above.
AphidI'd like to see a 72GB 384-bit card at the high end for <$2K, ideally around a 50% price premium on the 36GB version, which would still be a 50% margin on that added memory for AMD.
I'd love that as well, but I believe both of us know this is not really realistic.
Such product would likely be in the $3~4k range, which is honestly not bad still.

I have more hopes for a sub-$1k dual Arc B60 at the current point in time, 48GB of VRAM ain't bad at all at this price point.
GodisanAtheistBringing dual issue throughput on par with NV is a major area where AMD can improve its IPC. It's had dual issue since RDNA3, but the actual hit rate for dual issue ain't great.
The improvement on their dual issue on RDNA4 is really noticeable. On RDNA3 it was almost moot since it hardly ever came into effect, sadly.
Posted on Reply
Add your own comment
Nov 18th, 2026 10:44 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts