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

#151
Krit
olymind1And yes, i'm not going to buy NV because they're also expensive and/or slow-ish, the 5070 has only 12 GB vram and starts at 550€
RTX 5070 TI is worse than RTX 5070 what you get for the money 538€ vs 810€ that's 23-27% performance improvement for 51% more money. RTX 5070 is the best nvidia gpu by far in RTX 50 Series gpu line. RX 9070 XT also is better than RTX 5070 Ti p/p wise even though it's usually over msrp.
Posted on Reply
#152
medi01
john_AMD's latest financials do point to much more success in laptops.
Please elaborate. I haven't seen much of a split. But recall something about Radeon's being "strong".
Posted on Reply
#153
john_
medi01Please elaborate. I haven't seen much of a split. But recall something about Radeon's being "strong".
Well, clients going up means CPUs/APUs mostly, because of the low market share in GPUs and the fact that MSRP in AMD's RX 9000 wasn't remained for long to say that for months people where buying 9070s and 9060s. OEMs sill push much more AMD CPUs and GPUs anyway because they favor Nvidia in GPUs. And OEMs are those who make the big number. DIY market is single digit percentage, like 6-7% of the whole market. Laptops today is a bigger market than desktops, so probably a much higher number of laptop CPUs where sold than desktop CPUs, even considering that all the fuss is about desktop X3D chips. That's how I see it, doesn't mean it is correct. You might have a different opinion. Everyone else might have a different opinion.
Posted on Reply
#154
ARF
This will need to be revolutionary.
Today, the difference between AMD and Nvidia is astronomical - the new Radeon will need to be around at least 130% faster than the old Radeon RX 9070 that is just a lower-end mid-range, with respect to what is available as manufacturing nodes, and how many transistors can potentially be crammed onto the silicon.

Posted on Reply
#155
Krit
ARFThis will need to be revolutionary.
Today, the difference between AMD and Nvidia is astronomical - the new Radeon will need to be around at least 130% faster than the old Radeon RX 9070 that is just a lower-end mid-range, with respect to what is available as manufacturing nodes, and how many transistors can potentially be crammed onto the silicon.

RTX 5090 is way overpriced and overpowered (TDP wise) with useless 600w power consumption it's definitely not a typical gpu for an average user or gamer. It's actually first time in gpu history where manufacturer is selling a stove inside the pc. Usually 250w to 350w is the norm. As i understood AMD will relese smaller 384bit gpu with way less tdp than 600watts and performance may not hit RTX 5090 performance levels. To catch up RTX 5090 AMD shold relese gpu thats 91% faster than RX 9070 XT to mach average fps of RTX 5090. Can they do it at 350watts no they can't!!!! What they can do is something between RTX 4090 and RTX 5090 for better p/p and less absurd power consumption.

Yes RTX 5090 is fast if compare to AMD current offerings but when we look at price and power consumption it's complete trash with no efficiency at all. RTX 5080 is bad but RTX 5090 is way worse than that.

Right now cheapest RX 9070 XT is 660€ and cheapest RTX 5090 is 2400€ that's 91% faster for 264% more money and it consumes around 87% more power from wall.
Posted on Reply
#157
Krit
ARFThey can, of course they do. The key is to use TSMC's most advanced manufacturing node 2nm. RTX 5090 is an old technology, manufactured using the very dated 5nm process lithography ! ! :slap:
They can if they go flat out dumb as nvidia with high tdp like 600watts.
Posted on Reply
#158
ARF
KritThey can if they go flat out dumb as nvidia with high tdp like 600watts.
No, 2nm to 5nm is DOUBLE the transistor density. Up from the current 150 MTr/mm^2 to 300 MTr/mm^2.
Navi 48 is a medium sized silicon of only 357 mm^2 area.
They can make a UDNA x1 with 4 times the transistors of the poor Navi 48 - 200 billion transistors, GB202 has only 92.2 billion.
Posted on Reply
#159
ARF
KritYes but RDNA 4 is already 4nm not 5nm ?
No, it's just a marketing buzz word in order to mask / hide the stagnation and why they use one node for a half a decade already.
Posted on Reply
#160
TumbleGeorge
I'm going to contradict some of the recent opinions. AMD could very well catch up to the RTX 5090 in gaming with its next flagship. It will lag slightly in VRAM bandwidth, which can be fully compensated for with an increased and accelerated infinity cache. Whether it will happen depends entirely on the progress of GPU architecture and drivers. It will lag as before in Cuda-dependent applications.
Posted on Reply
#161
Krit
When looking at RX 7900 XTX results @4k against RTX 5090, 5090 is 74% faster.

Basically the new UDNA 384bit gpu should be around 74% faster than RX 7900 XTX to match RTX 5090 is it even possible probbably yes becouse RDNA4 skipped it's high end model.
Posted on Reply
#162
Dr. Dro
ARF:kookoo:
They can, of course they do. The key is to use TSMC's most advanced manufacturing node 2nm. RTX 5090 is an old technology, manufactured using the very dated 5nm process lithography ! ! :slap:

FYI, 5nm node is a 2018 node. 🤢🤮



en.wikipedia.org/wiki/5_nm_process

en.wikichip.org/wiki/5_nm_lithography_process
This is not the case. Nvidia's 4N is an optimized version of TSMC's 5 nm process. A very old article covers this:

www.tomshardware.com/features/nvidia-ada-lovelace-and-geforce-rtx-40-series-everything-we-know
KritWhen looking at RX 7900 XTX results @4k against RTX 5090, 5090 is 74% faster.

Basically the new UDNA 384bit gpu should be around 74% faster than RX 7900 XTX to match RTX 5090 is it even possible probbably yes becouse RDNA4 skipped it's high end model.
Matching a 5090 2 years late is pretty much a minimal requirement, unless the prices are like 9070XT's and positioned at the middle of Nvidia's stack. That isn't to say it won't be awesome if it does perform like a 5090 at like, 450 W instead of 600. But will be outshined by Rubin/Feynman for sure.
Posted on Reply
#163
alwayssts
I don't understand why it appears so outlandish for 96 CUs (what the uninformed call 6144sp, I call 12288sp) to compete with 4090.

4090 is what, 16384 @ 2730mhz (with a overclock up to ~3ghz) with 21gbps ram on a 384-bit bus.
This could conceivably be 12288sp @ 3.7-4ghz+...very similar to 4090 just with higher clock than units, higher memory speed to compensate smaller bus,
Clock guess is based on Apple N3B/N3E chips being 3.78/4.05ghz respectfully, and generally (over gens) those clocks transferring fairly close to GPUs.
If you line up the bandwidth (cache excluded, as we don't know it yet), it lines up as well. Might end up slightly faster overall with overclocking if 36gbps ram is true...but still a 4090 replacement (or slightly better). Given 4090 is a good card, but sometimes DLSS (especially transformer) can relegate it to <60fps (or sometimes 120fps, depending on game), something equal to slightly better would be a great card.

It would certainly be nice if it were something like half the price or so, and would fit the people that want "something better than a [next-gen] console" without perhaps breaking the bank.
Again, this wouldn't much surprise me if this were very-much competing with a '6080' with somewhat similar specs.

Appears pretty close to me, and probably the card I would want [assuming it was at least 24GB].
Similarly, with 144CUs (or so) the same would be true of the 170CUs of 5090; around 18% higher clock would be needed, and mem bw of ~37.33_gbps (with same cache).

Given 3nm is supposed to give ~18% (on n3e) or slightly more (~5%+ on n3p), and 40gbps ram is *possible* (or at least parity to 5090 with a 36gbps overclock), it all makes sense and lines up to me.

$1600-2000/~$1200-1500/$800-1000/$450-600/$300-400?

If this *were* to happen, this would be quite the departure for AMD, whom is VERY weary to release something above $1200, and actually cancelled their last attempt at that price (the 144CU RDNA4).

That said, MLID keeps implying the top-end die would be cut down to a 384-bit part (and perhaps it, and even the 256-bit part, may not exist as individual chips)...so maybe no super high-end for consumer.

Instead, a cut-down high-end chip and a 6080 competitor to entrap nVIDIA with better performance at lower price or higher perf for similar prices (given how the '80' series has trended).

Sounds good to me! I think people are thirsting for the 'old' high-end of ~$700-800; tangibly better than a new gen of consoles (a la ~4090?) but not beyond what many are willing to spend for such a device.
Posted on Reply
#164
Bagerklestyne
ARFThis will need to be revolutionary.
Today, the difference between AMD and Nvidia is astronomical - the new Radeon will need to be around at least 130% faster than the old Radeon RX 9070 that is just a lower-end mid-range, with respect to what is available as manufacturing nodes, and how many transistors can potentially be crammed onto the silicon.

You realise by asking for a card that's 130% faster than a current 9070, you're asking for a card that's 15% faster than a 5090.

They don't need to wallop 5090, just beat out the 5080 and probably 4090 but price it properly.
Posted on Reply
#165
ARF
BagerklestyneYou realise by asking for a card that's 130% faster than a current 9070, you're asking for a card that's 15% faster than a 5090.

They don't need to wallop 5090, just beat out the 5080 and probably 4090 but price it properly.
It will not sell. Remember why Ryzen became what it is today? Because it gave 8-core CPUs when the highest Intel had at the time was 4-core (Core i7-7700K) and 6-core (Core i7-8700K).
Remember that Nvidia will release RTX 6090 that will be at least 50-60% FASTER than the old RTX 5090.
If AMD doesn't match RTX 6090, it will be more market share loss.
Posted on Reply
#166
alwayssts
BagerklestyneYou realise by asking for a card that's 130% faster than a current 9070, you're asking for a card that's 15% faster than a 5090.

They don't need to wallop 5090, just beat out the 5080 and probably 4090 but price it properly.
Agree on relative card placement, but 'percentage placement' vs another card is a trickier thing to nail down given variability across certain games/features.

The thing I keep mentioning is to truly look at games. For the high-end one doesn't truly need something much faster than a 4090, but almost-certainly desire something faster than 5080 (or 9070xt).

I understand when I say that some will mention the evolution of features over time, and that's fair, but assuming featureset of RDNA5/UDNA is similar on PC as consoles, I think it's a fair place to desire.

Think of something like a 2080Ti, released end of 2018 (iirc?). A card like that more-or-less lasted, either at native or through use of DLSS, more-or-less this whole (20-26/28?) console generation (esp w/ 20% oc).

I think such a card (~4090+) would do the same for the next generation, especially if UDNA dramatically improves RT/PT (and/or lower-level 'performance'-level up-scaling) feature set...which it may. DLSS4 did.

I feel every generation of consoles is always preceded slightly earlier by a ~800-$1000 (or slightly less) card that gives PC users comfort they can play those games comfortably the whole length of the generation.

Some may argue for the current generation that was a 2080ti. I would argue that card for the future is not 5080 nor 9070xt, but will likely be their replacements (even if needed to be overclocked).

Something needs to finally be the improvement over 4080 that 5080 (purposely) was not, or rather bring ~4090 (with current featureset/RT/PT/FG/up-scaling perf) down to a lower price level, again unlike 5080.

As I've said before, 5090 truly is a gmmick of what can be done on a monolithic die and the margin they can achieve from those that would buy it. The vast majority do not need it (let-alone can they afford it).

I think most desire something like ~100TF FP32? So perhaps something like 6144sp (really 12288) dual-issue to 100TF using a clock of around 4ghz+?

Should be achievable within the clock capabilities of 3nm and 36gbps ram OC on a 256-bit bus...assuming there is sufficient cache...which I would imagine they will make sure there will be.

I would find something like this pretty exciting, given not only has the price of the '80' series gone out of control, they've been retisent to bring 4090(+) down to a lower price bracket.

If AMD were able to do both, that would be a major win IMHO.
ARFIt will not sell. Remember why Ryzen became what it is today? Because it gave 8-core CPUs when the highest Intel had at the time was 4-core (Core i7-7700K) and 6-core (Core i7-8700K).I
Remember that Nvidia will release RTX 6090 that will be at least 50-60% FASTER than the old RTX 5090.
If AMD doesn't match RTX 6090, it will be more market share loss.
Hard disagree. Most did not buy the 5090, and I don't see that changing with ever-increasing prices, even if it improves 'x'' amount.

Yes, *some* people will buy them, but as I've said before, I think most are looking for a card slightly faster than 4080/5080/9070xt, amd I think this is apparent given how incredibly well 9070xt has sold.

Up-scaling (current/future) Snowdrop to 4k60 with 'quality' up-scaling, maybe RT games with 'balanced' or 'performance' (liike Wukong or it's eventual upcoming peers with similar limitations).
Again, part of that is improving FSR to the point more-intensive levels of upscaling is as up to the task as possible (balanced or performance), the same ways DLSS4 has, which perhaps may come with Redstone [and/or with the next gen of cards]).
Posted on Reply
#167
Sol_Badguy
ARFIt will not sell. Remember why Ryzen became what it is today? Because it gave 8-core CPUs when the highest Intel had at the time was 4-core (Core i7-7700K) and 6-core (Core i7-8700K).
Remember that Nvidia will release RTX 6090 that will be at least 50-60% FASTER than the old RTX 5090.
If AMD doesn't match RTX 6090, it will be more market share loss.
That seems incredibly optimistic.
Latest review, 2160p:
www.techpowerup.com/review/msi-geforce-rtx-5080-expert-oc/31.html
5090 - 146%
4090 - 111%
5080 - 96%
4080 Super - 86 %
The 5090 is 52% faster than the 5080.
Assuming the 6080 will be halfway between 5080 and 5090, that means 128%.
So a 6090 being 52% faster than that means 195%.
Which then translates to 34% more than the 5090.
But you're hoping for much more than that.
At 2160p the 5090 is 32% faster than the 4090.
The 4090 itself is 29% faster than the 4080 Super.
And it was enough to be attractive, thus nVidia has this interval of 30-50% for the flagship 6090 to be faster than the 6080 in order to be attractive enough and justify the position in the product stack and corresponding price.

Personally I feel that 30% faster than the 5090 at 2160p is very realistic. 50-60% is just wildly optimistic. The performance gap will be so great that this time it will practically force nVidia to release something in-between, a cut-down 6090 of sorts, because this huge performance gap gives AMD a lot of room for maneuver and the realistic possibility to release something that can surpass the 6080. And something that slightly beats the 6080 will be priced the same at the most, given that AMD is not a market leader. So AMD won't need to compete at all at the $2000-2500 price bracket if it manages to compete successfully at the $1000-1200 price bracket.

AMD can make a midrange die (the 10070 XT), basically replicate the 9070 XT scenario for RDNA5, this will trade blows with the 6070 Ti which will be the cut-down midrange nVidia die.
For the 5090 and 6080 AMD can make a bigger die, the 10090 XT using the full die and the 10080 XT using a cut-down die.

I've said it before AMD doesn't need so much to counter the 6090 but the 5090 on the secondary market. The 6090 will sell regardless of what AMD does, even if AMD pulls a 6090 killer from its hat, the 6090 will sell even if only for the CUDA crowd. But for the gaming crowd sabotaging the 5090 reselling will be a very important battle that AMD needs to win. This is how they will gain market share, by convincing people that their new products are better than nVidia's previous generation (at the same price points, new vs used).
Posted on Reply
#168
TumbleGeorge
Sol_BadguyThe 5090 is 52% faster than the 5080.
56.25%.
Posted on Reply
#169
Krit
@4k it's 53% faster but it costs 130% more. :kookoo:
Posted on Reply
#170
john_
ARFThey can make a UDNA x1 with 4 times the transistors of the poor Navi 48 - 200 billion transistors, GB202 has only 92.2 billion.
So, they will make a GPU with double the transistor count of Nvidia's GPU, at a much more expensive node and try to sell it to people who will find any excuse possible to NOT buy AMD. While trying to maintain >40% profit margin, when Nvidia sells with >70% profit margin, meaning Nvidia has much more room to lower prices and still have a healthy profit margin. And in a market that game developers prefer using proprietary Nvidia libraries, instead of GPU agnostic libraries that are optimized and compatible to run with any vendor GPU.

That's a nice plan to bankrupt a company.
Put your usual 🤣 smile. I don't expect a reply.
Posted on Reply
#171
medi01
Dr. DroMatching a 5090 2 years late is pretty much a minimal requirement
But why?

Note the gap between xx80 card and 5090.
It has... grown a lot.

AMD has no acute need to fight beyond xx80.
6080 is also unlikely to beat 5090, given 5080 did not beat 4090.
Posted on Reply
#172
Dr. Dro
medi01But why?

Note the gap between xx80 card and 5090.
It has... grown a lot.

AMD has no acute need to fight beyond xx80.
6080 is also unlikely to beat 5090, given 5080 did not beat 4090.
It's not rocket science, you're either building and selling flagship class hardware or affordable pack-a-punch midrangers like the 9070 XT, which have a low profit margin, despite the volume.

This strategy works for building market share within a certain segment, the problem with following the latter strategy long-term inevitably relegates the coveted position of premium vendor and after some time you become the de facto cheap option.

It's not exactly a perfect outlook if your latest and greatest technology just about caught up to a half-decade-old flagship from your competitors.
Posted on Reply
#173
Krit
Dr. DroIt's not rocket science, you're either building and selling flagship class hardware or affordable pack-a-punch midrangers like the 9070 XT, which have a low profit margin, despite the volume.
RX 9070 XT has a good profit margin definitely better than RX 7800 XT. RX 9070 XT has very similar specs and die size but it costs closer to 200€ more. It's just that RTX 5080 or way worse RTX 5090 is complete rip off. Paying for god knows who mostly the empty air.

For example RTX 2080Ti, GTX 1080Ti or GTX 980Ti was scaling way better with RTX 2070, GTX 1070 and GTX 970 now we are paying just for empty air if you want a **80Ti tier gpu or even just a **80 higher mid range gpu.

RTX 5090 is 350% more expensive than RTX 5070
RTX 2080Ti was 100% more expensive than RTX 2070
GTX 1080Ti was 67% more expensive than GTX 1070
GTX 980Ti was 89% more expensive than GTX 970

That's a night and day difference or very clear rip off vs normal margin. RX 9070 XT and RTX 5070 is a normal to higher margin and definitely not low profit margin.
If anything has low profit margin than its intel ARC B570/B580 look at specs and prices. :)

Those people who buy these pointless gpus RTX 5080, RTX 5090 and RTX 5060 Ti 16GB are supporting these stupid dumb high prices by default. And over time it will get only worse because nvidia already tasted the blood and there is no way back until someone kicks hard in nvidia's teeth and delivers better product for lower price. May hope on UDNA but RDNA4 already is overpriced with higher profit margins than RDNA3 which aviously is not a good sign at all. I wouldn't be surprised at all if there was some kind of agreement between these companies to not lower prices below a certain threshold for better milking capabilities and i'm not talking about intel aviously.
Posted on Reply
#174
Dr. Dro
KritRX 9070 XT has a good profit margin definitely better than RX 7800 XT. RX 9070 XT has very similar specs and die size but it costs closer to 200€ more. It's just that RTX 5080 or way worse RTX 5090 is complete rip off. Paying for god knows who mostly the empty air.

For example RTX 2080Ti, GTX 1080Ti or GTX 980Ti was scaling way better with RTX 2070, GTX 1070 and GTX 970 now we are paying just for empty air if you want a **80Ti tier gpu or even just a **80 higher mid range gpu.

RTX 5090 is 350% more expensive than RTX 5070
RTX 2080Ti was 100% more expensive than RTX 2070
GTX 1080Ti was 67% more expensive than GTX 1070
GTX 980Ti was 89% more expensive than GTX 970

That's a night and day difference or very clear rip off vs normal margin. RX 9070 XT and RTX 5070 is a normal to higher margin and definitely not low profit margin.
If anything has low profit margin than its intel ARC B570/B580 look at specs and prices. :)

Those people who buy these pointless gpus RTX 5080, RTX 5090 and RTX 5060 Ti 16GB are supporting these stupid dumb high prices by default. And over time it will get only worse because nvidia already tasted the blood and there is no way back until someone kicks hard in nvidia's teeth and delivers better product for lower price. May hope on UDNA but RDNA4 already is overpriced with higher profit margins than RDNA3 which abiously is not a good sign at all. I wouldn't be surprised at all if there was some kind of agreement between these companies to not lower prices below a certain threshold for better milking capabilities and i'm not talking about intel aviously.
Overly simplistic view of an exceptionally complex situation. Market positioning isn't simple, and if done incorrectly, it can and will damage a product line. Margins must account for:
  • 8-10+ years of driver updates
  • Research and development of features for the hardware that will be released for free
  • Research and development of future hardware architectures
  • Product branding and design costs
  • Market research and product distribution costs
  • RMA costs for any eventual defective units across AIBs (since AIBs purchase assembly kits per card)
  • All while retaining a pricing balance that the market will bear
Short-term, lowering your margins and eating the costs of some of the things described above can be a strategy to build and regain both market and mind share. This is the strategy AMD has used with the RX 480, 5700 XT and 9070 XT, aggressively undercutting the competition since they couldn't bring the fight to them in the technology and performance fronts. All three of these products were commercially successful - unsurprisingly, people like cheap things.

Long-term, it means losing the coveted status of premium vendor and poorly supported, underdeveloped products, and this has resonated in the real world with AMD's exceptionally poor support of some of their earlier generation architectures. Example: GTX 900 series just about to enter EOL this October, after 11 years of driver updates, their competing R9 300 and Fury series products have been discontinued for years. Polaris and Vega GPUs have been on a long term support lifeline and haven't been receiving updates to their codebase in a long time as well, leading to compatibility issues with newer titles (Battlefield 6 doesn't work on Vega without spoofing minimum driver version, etc.)

Arc's situation is a can of worms entirely its own. You may have noticed that they have scaled back significantly on Arc lately, B770 is still vaporware and unlikely to release at this point, and that there are few units of B570/B580 GPUs available to purchase. They're playing it very safe, and Intel's upper management hasn't shied away from passing the message that they're a CPU first company, using their wafer allocation from TSMC to make more CPUs than Arc cards.

Besides, it won't be a mediocre product that will "kick Nvidia's teeth", you call a RTX 5090 a "pointless GPU" while simultaneously debating how will AMD's 2027 product tackle its 2022 predecessor and perhaps have an honest shot at it years down the road, and expect "someone to kick Nvidia's teeth in". Have you noticed the cognitive dissonance here?
Posted on Reply
#175
medi01
Dr. DroIt's not rocket science, you're either building and selling flagship class hardware or affordable pack-a-punch midrangers like the 9070 XT, which have a low profit margin, despite the volume.
Are you telling me that 5080 and below GPUs by the Filthy Green bring little to no monez?

Not note that x900 series GPU development also costs money.
And it is a very VERY specific section of the market, where being just a little below the competitor is DEADLY.
It is for the people who want "the best, DGAF how expensive".

Questionable reward, high risk.
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