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

[igormp]

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

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.

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.

For sure gamers would love that, too bad they don't pay enough to have it
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.

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.

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.

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.

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.

 

[Dr. Dro]

-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.

Hmm, you do have a point. This is probably why the SP metric has fallen in disuse. 6144 (96 CU, 48 WGP) is doable, 12288... isn't

 

[Zendou]

That's what I am saying. Always finding an excuse to NOT buy an AMD GPU. It boggles the mind that people find insulting business practice X when AMD is doing it, and justified when Intel/Nvidia is doing it.

No, no, no. AMD should stop offering gifts. Prices close to Nvidia, features close to Nvidia and every new feature exclusive to new cards. After 3-5-10 years people might... MIGHT stop pointing the finger at AMD, really I mean REALLY start questioning the practices of the 4 trillion company with the 75% profit margin that controls over 80% of the gaming market and influences developers, tech press and even big AIBs and start doing something positive with their wallets. Until today they are just trying to find any excuse to justify using their wallet to help Nvidia's monopoly.

So AMD should run their GPU division into the ground on principal and if they somehow exist after that, boy won't the public be embarrassed? This seems to ignore any and all economic principal's and realities. Nvidia is currently winning because they make a better product than the competitors. Very akin to the 14nm++++++++ days with intel, where they did not challenge them for so long that Intel got complacent. Now did AMD price near intel and wait for the public to admonish Intel and that led them to victory? Oh, they made a better product at a better price and you could measure it objectively to prove it. Yeah AMD should do that with their GPUs.

 

[john_]

So AMD should run their GPU division into the ground on principal and if they somehow exist after that, boy won't the public be embarrassed? This seems to ignore any and all economic principal's and realities. Nvidia is currently winning because they make a better product than the competitors. Very akin to the 14nm++++++++ days with intel, where they did not challenge them for so long that Intel got complacent. Now did AMD price near intel and wait for the public to admonish Intel and that led them to victory? Oh, they made a better product at a better price and you could measure it objectively to prove it. Yeah AMD should do that with their GPUs.

For at least the last 15 years people wanting from AMD to
- offer much cheaper options
- be almost as good as Intel's and Nvidia's to force those companies to lower prices
- but not be equal or better than Intel's or Nvidia's, because they still wanted to feel that buying Nvidia or Intel hardware puts them on the top class, gives them the right to laugh at those "poor people buying AMD hardware"

AMD tried their best all those years to get market share and support. They managed to do that in CPUs, not just because they where offering better CPUs, but because Intel was going from one disaster to the next. Intel was pushing 300W to win in gaming and people still preferred Intel. They failed in GPUs. Not because they had much worst products, but because the average online person was too friendly to Nvidia, too hostile to AMD. The last 15 years people keep finding excuses to NOT buy AMD hardware. They keep playing the same song "Make good enough products, so I can buy Nvidia, because I better cut my hand than touching an AMD GPU".
In many cases AMD was offering better products at better pricing. But people where always trying to find an excuse to pay more to Nvidia to get less. RX 6600 vs RTX 3050 is a major example. RTX 3050 more expensive, much slower, it probably sold 10 times more than RX 6600. There will always be an excuse. And every excuse will be "HUGE" reason to not buy AMD.

 

[blacksea76]

Without knowing pricing, i have no interest in it. If it's the same strategy like RDNA4 it will have the same success, none.

No success with 9070 series? I was watching an online seller here in the UK for ages after it launched, and it sold like fresh bread. Where do you people get the data?

 

[mkppo]

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.

Mate, you give Visible Noise an answer that actually makes sense which he won't have a reply to and he will slowly chicken out of the thread and not reply. It's what he does.

What an absolutely stupid line to say raster is irrelevant. Then doubles down to ask what games are ROP limited like he knows something. Utter BS

I think he put me on his ignore list after being schooled a few times. It's just how it is. I also see he has a fan named wolf.

Good riddance I guess.

 

[Bomby569]

No success with 9070 series? I was watching an online seller here in the UK for ages after it launched, and it sold like fresh bread. Where do you people get the data?

everybody is buying nvidia cards as usual. You're all delusional as usual.

 

[calhau]

No success with 9070 series? I was watching an online seller here in the UK for ages after it launched, and it sold like fresh bread. Where do you people get the data?

Q1 2025 market analysis, 92% Nvidia, 8% AMD.

And things got worse since Q1, the 100$ to subsidise the fake MSRP is gone, and 5070's stock is now available closer to MSRP, making the 9070's less attractive.
From what I've seen (this is for my country might be different in others) the 9070s price are much more inflated than that of the 5070s over announced MSRP, so they sell a lot less.
For example the cheapest 9070xt in my country now is 720€ (should find one at 630€ if 600$ was real), the cheapest 5070ti is 812€ (should be 789€), if you normalise price/performance (using TPU's relative performance values in raster) the value of the ti is 767€ compared to the xt 720, that's just 46€ extra for a card with better upscaling, ray tracing, energy efficiency and that is miles better in some common productivity cases.

If AMD MSRP wasn't fake they would be selling, but it was, and with the gains of the Euro Vs dollar, here old stock is creating problems, because sellers bought it at a higher price (no 100 bucks discount and 13 cents less per € than now) and they don't want to cut their profits so the price continues inflated, basically using it as an upsale for the 5070ti that gives them more profit than if they sold a 9070xt.

 

[Denver]

AMD doesn't need 150-200 CUs. 9070XT is performing more or less like a 7900XTX based on TPUs ranking(9070XT 6% slower). So a 9000 series with 96 CUs will be probably between 5080 and 4090. With 128 CUs it will be at 5090 level of performance.

A 96CU RDNA 4 GPU would outperform the RTX 4090, and even the 5090 in some scenarios (example below), despite using GDDR6.
Previous projections for UDNA seem accurate, around a 20% performance increase per CU compared to RDNA4.
The strong emphasis on doubling AI and RT performance likely stems from Sony’s partial funding of the architecture’s R&D, driven by their demand for significant RT and ML advancements.
As a result, a 96CU UDNA could deliver 10–20% higher rasterization performance than the 5090, with substantially greater gains in ray tracing.
But I still expect flagship UDNA N58/N51 to have 2x more CU than N48.

 

[medi01]

Make OEM escape Filthy Green's grip and start doing AMD dGPUs again. Typing this from AMD Advantage G15 with 6800M.

cool, can we get something more than midrange please?
150-200 CUs please.

Is 5080 midrange?

The only card that is really out of reach or AMD is the xx90, and f*ck that shit anyhow.

 

[calhau]

A 96CU RDNA 4 GPU would outperform the RTX 4090, and even the 5090 in some scenarios (example below), despite using GDDR6.
Previous projections for UDNA seem accurate, around a 20% performance increase per CU compared to RDNA4.
The strong emphasis on doubling AI and RT performance likely stems from Sony’s partial funding of the architecture’s R&D, driven by their demand for significant RT and ML advancements.
As a result, a 96CU UDNA could deliver 10–20% higher rasterization performance than the 5090, with substantially greater gains in ray tracing.
But I still expect flagship UDNA N58/N51 to have 2x more CU than N48.

It would be 36% slower at 4k raster on average than a 5090ti according to your estimations.

Alan Wake 2 why not Elden Ring as your comparison? it's overall a much more popular game.
Individual benchmarks are very important, but it's up to individuals to judge their value based around their use case, not to make any form of argument for which card is the fastest.

What's the point of biased speculation? Gaining points with AMD fanboys? Misleading less aware ppl so they are more likely to buy AMD?

Why do ppl are so attached to brand that they make things worst for others and eventually them?

 

[GodisanAtheist]

View attachment 409054
View attachment 409069
It would be 36% slower at 4k raster on average than a 5090ti according to your estimations.

Alan Wake 2 why not Elden Ring as your comparison? it's overall a much more popular game.
Individual benchmarks are very important, but it's up to individuals to judge their value based around their use case, not to make any form of argument for which card is the fastest.

What's the point of biased speculation? Gaining points with AMD fanboys? Misleading less aware ppl so they are more likely to buy AMD?

Why do ppl are so attached to brand that they make things worst for others and eventually them?

- Yep. Thing is AMD is not beating NV on performance per area. They're nearly even with RDNA4 on raster and still slightly but significantly behind in RT. So a 500mm2 RDNA4 die is not going to out perform a 750mm2 Nvidia part.

But AMD doesn't have to. AMD doesn't need to take the Halo part, NV left a dumptruck sized performance and die size gap between the $1200 5080 and the $2500 5090.

A 500mm2 AMD part that splits the difference for $1500 would net them a win by most practical buyers standards.

 

[Denver]

View attachment 409054
View attachment 409069
It would be 36% slower at 4k raster on average than a 5090ti according to your estimations.

Alan Wake 2 why not Elden Ring as your comparison? it's overall a much more popular game.
Individual benchmarks are very important, but it's up to individuals to judge their value based around their use case, not to make any form of argument for which card is the fastest.

What's the point of biased speculation? Gaining points with AMD fanboys? Misleading less aware ppl so they are more likely to buy AMD?

Why do ppl are so attached to brand that they make things worst for others and eventually them?

Alan Wake 2 is well optimized for Nvidia and AMD and is GPU bound, it's even a game used by Nvidia to promote RT.
Elden Ring is a light and buggy game that runs well on anything.

My argument is less biased than yours or the TPU test list which uses games known to be broken or extremely lightweight like CS2, at the end of the day, this average is in no way representative of the average performance of GPU-bound AAA titles.

I don't understand why you're crying. Beware that people who buy $3k GPUs will stop buying because of my estimation of a product that doesn't even exist.

 

[Visible Noise]

in about four years when AMD figures out how raytrace triangles in any sort of a competitive way, the industry will have moved on to voxel based rendering with Restir DI/GI.

AMD, always late to the party.

 

[shadyhero]

Unless it's some earth-shattering improvement, I'm afraid I'll probably still go with NVidia. The pure raster performance is not that relevant anymore. Subsequently, the power of the RTX/DLSS ecosystem and its market penetration are just impossible to ignore and worth the premium for me. But I guess it'll be a good option for people who have time for tweaking stuff and playing with tools such as Optiscaler.

that's if 60 series is actually any good. given nvidia's current track record they'll probably find a way to mess it up, be it price or die shrinkflation or a combination of the two.

 

[ARF]

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

Meanwhile GB202 has "only" 24.5k shaders lol.

If AMD wants to succeed, it really needs to redesign its architecture and make UDNA x1 with at least 30k shaders.

 

[GodisanAtheist]

Meanwhile GB202 has "only" 24.5k shaders lol.

If AMD wants to succeed, it really needs to redesign its architecture and make UDNA x1 with at least 30k shaders.

- Nah, these massive chips come with a bunch of overhead. The 5080 is a 378mm^2 die that is only 34% slower than the 750mm^2 5090 at 4K (Said another way, the 5090 is 52% faster than the 5080 at 4K).

Even if the 5090 was fully unlocked, the massive drop off in performance for the massive increase in die size is a purely Nvidia big dog move.

If AMD made a 500MM^2 chip, but built it to clock to the moon with looser tolerances for heat and power consumption, they'd get most of the way there without nearly the same manufacturing cost or opportunity cost in Ryzen chips.

 

[ARF]

^^^^ RTX 5090 is so fast (and thus, faster than RTX 5080), that it needs a completely new CPU architecture in order not to be sandbagged there.

 

[GodisanAtheist]

^^^^ RTX 5090 is so fast (and thus, faster than RTX 5080), that it needs a completely new CPU architecture in order not to be sandbagged there.

-Yeah but that's an unbalanced design. Why build a $2500 behemoth "gaming" chip that cannot even reach it's full potential today.

Why not build a $1500 big die that can reach it's full potential today (and maybe fall 10-15%) behind the 5090.

That's AMD's bread and butter, doing as much as possible with as little as possible.

 

[Dr. Dro]

-Yeah but that's an unbalanced design. Why build a $2500 behemoth "gaming" chip that cannot even reach it's full potential today.

Why not build a $1500 big die that can reach it's full potential today (and maybe fall 10-15%) behind the 5090.

That's AMD's bread and butter, doing as much as possible with as little as possible.

That'd be the RTX 5080, but to be entirely honest this was also an issue the RTX 4090 faced and continues to face, the problem is that the CPU side of tech is completely stalled right now. With Arrow Lake being generally a performance regression, and Raptor Lake being just a cache boosted version of Alder Lake, an architecture that is now as old as Zen 3 itself... Intel hasn't brought anything noteworthy to the table in the first half of this decade, and growth has been glacial over at AMD's side too. Zen 3 CPUs continue to linger at the low end, AM4 a socket that seemingly refuses to go. Zen 4 matched Intel chips and Zen 5 is a nominal upgrade at best in the desktop side, especially for gaming, with most gains realized at the server and workstation segment thanks to improved AVX-512 support.

 

[Makaveli]

That'd be the RTX 5080, but to be entirely honest this was also an issue the RTX 4090 faced and continues to face, the problem is that the CPU side of tech is completely stalled right now. With Arrow Lake being generally a performance regression, and Raptor Lake being just a cache boosted version of Alder Lake, an architecture that is now as old as Zen 3 itself... Intel hasn't brought anything noteworthy to the table in the first half of this decade, and growth has been glacial over at AMD's side too. Zen 3 CPUs continue to linger at the low end, AM4 a socket that seemingly refuses to go. Zen 4 matched Intel chips and Zen 5 is a nominal upgrade at best in the desktop side, especially for gaming, with most gains realized at the server and workstation segment thanks to improved AVX-512 support.

100% agreed.

This is a good take on whats going on now on the cpu side.

I'm looking forward to seeing a scaling review on Zen 4,5,6 with a 4090 and 5090 sometime in the future.

And if NL isn't crap add that in there too.

 

[Nhonho]

AMD must understand (the obvious) that people want maximum image quality from GPU-encoded videos. AMD should implement the highest video image quality preset for each codec in its GPUs, in order to its GPUs can encode videos with quality as good as videos encoded by CPUs at maximum quality preset, and preferably in two passes. AMD must realize that people don't use CPUs to encode videos anymore and that GPUs are no longer just used for games or 3D apps.

 

[ARF]

AMD must understand (the obvious) that people want maximum image quality from GPU-encoded videos. AMD should implement the highest video image quality preset for each codec in its GPUs, in order to its GPUs can encode videos with quality as good as videos encoded by CPUs at maximum quality preset, and preferably in two passes. AMD must realize that people don't use CPUs to encode videos anymore and that GPUs are no longer just used for games or 3D apps.

Of course, GPU acceleration is faster, better and more efficient.
AMD must understand that it needs to COMPETE in the graphics market, something that they forgot to do many years ago.

 

[Frick]

The heck are you talking about? When was the last time amd offered any gifts? 15 years ago? Starting from 2015 - tell me what generation was amd offering "gifts"?

I presume you mean bundles. Starfield comes to mind, which was included with even lower end stuff.

Of course, GPU acceleration is faster, better and more efficient.
AMD must understand that it needs to COMPETE in the graphics market, something that they forgot to do many years ago.

That's one of the goals with UDNA, isn't it?

 

[Zazigalka]

Its the same crap at nvidia camp , you have to twidle around to get dlss 4 since older games dont have it. I cant be bothered doing that.

The main issue with fsr 4 is that because of how few cards actually support it, its not going to get traction. Now imagine UDNA launching with yeyt a new version of fsr that is exclusive to the new gpus. No thanks, id stick to nvidia.

You have to select preset K in the nvidia app, all there is to do to get the transformer model.